Helsinki University of Technology Department of...
Transcript of Helsinki University of Technology Department of...
Helsinki University of Technology Department of Industrial Engineering and Management Doctoral Dissertation Series 2009/12 Espoo 2009
BIDDER SUPPORT IN ITERATIVE, MULTIPLE-UNIT COMBINATORIAL
AUCTIONS
Riikka-Leena Leskelä Dissertation for the degree of Doctor of Science in Technology to be presented with due permission of the Department of Industrial Engineering and Management for public examination and debate in auditorium TU2 at Helsinki University of Technology, Espoo, Finland, November 14
th, 2009 at 12 noon.
Helsinki University of Technology Department of Industrial Engineering and Management
Helsinki University of Technology Department of Industrial Engineering and Management
P.O. Box 5500 FI-02015 TKK, Finland
Tel. +358 9 451 2846 Fax +358 9 451 3665 Email: [email protected]
Internet: http://www.tuta.tkk.fi/en
© Riikka-Leena Leskelä
ISBN 978-952-248-127-6 (print) ISBN 978-952-248-128-3 (electronic)
ISSN 1797-2507 (print) ISSN 1797-2515 (electronic)
http://lib.tkk.fi/Diss/2009/isbn9789522481283/
All rights reserved. No part of this publication may be reproduced, stored in retrieval
systems, or transmitted in any form or by any means, electronic, mechanical, photocopying, microfilming, recording, or otherwise without permission in writing from
the publisher. Cover photos by Mike Lietz (right) and alias psiconauta (left) through the Flickr service
under the Creative Commons license.
Yliopistopaino Helsinki
AB
ABSTRACT OF DOCTORAL DISSERTATION HELSINKI UNIVERSITY OF TECHNOLOGY P.O. BOX 1000, FI-02015 TKK http://www.tkk.fi
Author Riikka-Leena Leskelä
Name of the dissertation Bidder Support in Iterative, Multiple-Unit Combinatorial Auctions
Manuscript submitted 15.5.2009 Manuscript revised 23.9.2009
Date of the defence 14.11.2009
Faculty
Department
Field of research
Opponent(s)
Supervisor
Instructors
Information and Natural Sciences
Industrial Engineering and Management
Decision and Negotiation Support
Prof. Wojtek Michalowski, University of Ottawa
Prof. Hannele Wallenius, Helsinki University of Technology
Prof. Hannele Wallenius, Helsinki University of Technology
Prof. Jyrki Wallenius, Helsinki School of Economics
Abstract
This thesis is about supporting the bidders’ decision making in iterative combinatorial auctions. A combinatorial auction refers to an auction with multiple (heterogeneous) items, in which bidders can submit bids on packages. Combinatorial auctions are challenging decision making environments for bidders, which hinders the adoption of combinatorial mechanisms into practice. Bidding is especially challenging in sealed-bid auctions. Bidders do not know the contents of other bidders’ bids and hence cannot place bids that would team up with existing bids to become winners. The objective of this study is to develop and test support tools for bidders in semi-sealed-bid, iterative combinatorial auctions. The tools are designed for reverse auctions, but can easily be applied to a forward setting.
The Quantity Support Mechanism (QSM) is a support tool, which provides the bidders with a list of bid suggestions. The bid suggestions are such that if submitted, they would become provisional winners. The QSM benefits both bidders and the buyer, because it chooses suggestions that are most profitable for the bidders while decreasing the total cost to the buyer. The QSM is based on a mixed integer programming problem.
The QSM was tested in two simulation studies. The results of the studies indicated that the QSM works well – it is much better to use the QSM than no support – but that it does not necessarily guide the auctions to the efficient allocation. The QSM was also integrated into an online auctions system, and tested with human subjects. The results of the laboratory experiment showed that the performance of the QSM is dependent on the bidders’ behavior and the kind of bids they place in the auction. The user interface of the auction was good. I also observed bidders’ strategies, and could identify different bidder types corresponding to those reported in earlier studies. The experiment also showed the importance of experience in complex bidding environments.
The simulation studies and the laboratory experiment showed that the QSM is too dependent on the existing bids in the bid stream, which causes the auctions to end in inefficient allocations. In order to overcome this problem we designed another support tool, the Group Support Mechanism (GSM). The main logic in the GSM is similar to the QSM. The main difference is that instead of solving for one bid that complements existing bids to become a winner, the GSM can suggest several bids for different bidders. Together this set of bids would then become provisionally winning. The preliminary tests show significant improvement in the efficiency of the auction outcomes when the GSM was used instead of the QSM.
Future research includes the further development of the GSM and its testing with simulations and human subjects. Also, bidder behavior, bidder strategies and the effect of learning and experience in combinatorial auctions should be further studied. This is important because bidders’ behavior in the auctions affects the auction design and the requirements for the user interface.
Keywords Auctions, Combinatorial Auctions, Bidder Support, Online Auctions, E-auctions
VÄITÖSKIRJAN TIIVISTELMÄ TEKNILLINEN KORKEAKOULU PL 1000, 02015 TKK http://www.tkk.fi
Tekijä Riikka-Leena Leskelä
Väitöskirjan nimi Bidder Support in Iterative, Multiple-Unit Combinatorial Auctions
Käsikirjoituksen päivämäärä 15.5.2009 Korjatun käsikirjoituksen päivämäärä 23.9.2009
Väitöstilaisuuden ajankohta 14.11.2009
Tiedekunta
Laitos
Tutkimusala
Vastaväittäjä(t)
Työn valvoja
Työn ohjaaja
Informaatio- ja luonnontieteiden tiedekunta
Tuotantotalouden laitos
Päätöksenteon ja neuvotteluiden tukeminen
Prof. Wojtek Michalowski, University of Ottawa
Prof. Hannele Wallenius, TKK
Prof. Hannele Wallenius, TKK, Prof. Jyrki Wallenius, HSE
Tiivistelmä Tämä tutkimus koskee tarjoajien päätöksenteon tukemista kombinatorisissa huutokaupoissa. Tarjousten
tekeminen kombinatorisissa huutokaupoissa on haastavaa – etenkin suljetuissa huutokaupoissa. Voittavat tarjoukset täydentävät toisiaan; niiden summa on kaupan kohteena oleva hyödykekombinaatio. Suljetussa huutokaupassa tarjouksen tekijät eivät kuitenkaan tiedä toistensa tarjousten sisältöä, joten he eivät osaa tehdä tarjouksia, jotka täydentäisivät muita tarjouksia. Tutkimuksemme tarkoituksena on kehittää ja testata työkaluja tarjouksen tekijöille puolisuljettuihin, iteratiivisiin kombinatorisiin huutokauppoihin.
Kehittämämme työkalu, Quantity Support Mechanism (QSM), ehdottaa tarjoajille tarjouksia joista mikä tahansa olisi kyseisellä hetkellä voittajien joukossa. Tarjoajan tehtäväksi jää päättää, haluaako hän tehdä jonkin ehdotetuista tarjouksista. QSM hyödyttää molempia osapuolia, sillä sen tekemät ehdotukset ovat tarjoajille mahdollisimman voitollisia ja samalla vähentävät ostajan kokonaiskustannuksia (kun kyseessä on käänteinen huutokauppa). QSM pohjalla on kokonaislukuoptimointitehtävä.
QSM:ia testattiin simuloimalla. Simulointien tulokset osoittivat, että QSM toimii hyvin – on parempi käyttää QSM:ia kuin olla ilman tukea – mutta sen käyttö ei aina takaa, että huutokauppa päättyy tehokkaaseen allokaatioon. QSM myös integroitiin osaksi Internet-pohjaista huutokauppajärjestelmää. Tämä mahdollisti QSM:n testaamisen koehenkilöillä. Kokeen tulokset osoittivat, että QSM:n toimiminen riippuu siitä, millaisia tarjouksia tarjoajat ovat huutokaupassa tehneet. Huutokaupan käyttöliittymä todettiin toimivaksi. Tutkin myös koehenkilöiden käyttämiä strategioita ja tunnistin niiden joukosta samantyyppisiä strategioita kuin aikaisemmissa tutkimuksissa tarjoajien on havaittu käyttävän. Koe osoitti myös kokemuksen tärkeyden monimutkaisissa huutokaupoissa.
Simulaatiot ja koehenkilöillä tehty testaus osoittivat että QSM on liian riippuvainen olemassa olevista tarjouksista. Tästä seuraa mm. että huutokaupat eivät pääty tehokkaaseen allokaatioon. Ratkaistaksemme tämän ongelman kehitimme toisen tukityökalun, the Group Support Mechanismin (GSM). GSM toimii pääpiirteissään samalla tavalla kuin QSM. Suurin ero on, että GSM ehdottaa tarjouksia yhtä aikaa useammalle tarjoajalle. Yhdessä nämä kaikki tarjoukset pääsisivät voittajien joukkoon, mutta eivät yksinään kuten QSM:n ehdotukset. Alustavat testit osoittavat GSM:n parantavan huomattavasti huutokauppojen lopputulosten tehokkuutta QSM:iin nähden.
Jatkossa keskitymme kehittämään GSM:ia ja testaamme sitä simuloinneilla ja koehenkilöillä. Tarjoajien käyttäytymistä, strategioita ja erityisesti oppimisen ja kokemuksen karttumisen vaikutusta käyttäytymiseen tulisi myös tutkia lisää. Tämä on tärkeää, sillä tarjoajien käyttäytyminen vaikuttaa huutokaupan suunnitteluun ja käyttöliittymältä vaadittaviin ominaisuuksiin.
Asiasanat: huutokaupat, kombinatoriset huutokaupat, päätöksenteon tukeminen, Internet-huutokaupat
AB
Acknowledgements
Just like no man is an island, no thesis is written without help and support from other people. Thus, now that I have completed my thesis, I want to express my gratitude to the people, who have either directly or indirectly influenced this process during the past years.
First and foremost I want to thank my advisors, Professors Hannele and Jyrki Wallenius. Besides introducing me to the interesting world of auction research, you have been a constant support throughout my PhD studies. You always had time for me, and you always, always had faith in me – even when I lost faith in myself. I also want to thank my co-authors, M.Sc. Valtteri Ervasti, whose programming skills made many of my ideas possible, Professor Murat Köksalan, with whom I have had several fruitful discussions over the years, and Professor Jeffrey Teich. It has been a privilege to work with you.
I am also grateful to the two pre-examiners of my thesis, Professor David Olson and Professor Karla Hoffman, for finding time in their busy schedules to get acquainted with my work and for giving comments that helped me improve my thesis. I also want to thank everybody in the Graduate School in Systems Analysis, Decision Making and Risk Management for creating an intellectually stimulating environment.
I want to thank the Emil Aaltonen foundation, the Kemira Oyj foundation, and the KAUTE foundation for financial support.
I have enjoyed the past five years of my life, and for that I want to thank my colleagues, my friends and my family. The faculty and staff of the Institute of Strategy have made this a great place to work at. Thanks to you and our 2:00 pm coffee (tea) breaks it is really nice to come to work every day. As for my friends and sisters, thank you for all the wonderful times we have shared together; all the laughs and tears, the singing and the travels, and all the great memories. You made sure I actually had a life outside work! I especially want to thank Johanna, without whom I would not have become excited about Economics, and without whom I would not be here today. I also want to thank my piano teacher Hanne for showing me that I too can be creative, and for giving me confidence to use my creativity.
Last but not least I want to thank my mother and my father. You taught me the value of education with your example, and you have been there for me every step of the way in all the ways that count.
Otaniemi, October 2009,
Riikka-Leena Leskelä
i
TABLE OF CONTENTS
I INTRODUCTION...........................................................................................1
1 DESIGNING AUCTIONS...............................................................................6
1.1 Desirable Properties of Auctions...................................................................6
1.2 Market Settings..............................................................................................7
1.3 Auction Mechanisms ..................................................................................10
II LITERATURE.............................................................................................13
2 THE ORIGINS OF AUCTION THEORY: SINGLE-ITEM AUCTIONS ...........13
2.1 The Independent Private Values Model ....................................................14
2.1.1 Strategic Equivalence of Auction Mechanisms in the IPV Model...................14 2.1.2 The Revenue Equivalence Theorem................................................................17
2.2 The Common Value Model .......................................................................18
2.3 The Affiliated Values Model.......................................................................19
2.4 The Popular English Auction .....................................................................21
2.5 Optimal Auctions ........................................................................................23
2.6 Empirical Studies of Auctions ....................................................................25
2.6.1 Field Studies ......................................................................................................25 2.6.2 Experimental Studies.........................................................................................25
2.7 Criticism of the Traditional Single-Unit Models .......................................27
2.8 Multiple-Unit Auctions...............................................................................28
3 MULTIPLE-ITEM AUCTIONS – AN INTERDISCIPLINARY FIELD ..........31
3.1 Non-Combinatorial Auction Mechanisms .................................................33
3.1.1 Sequential Auctions...........................................................................................33 3.1.2 Simultaneous Auctions......................................................................................33
3.2 Combinatorial Auctions..............................................................................36
3.2.1 Challenges with Combinatorial Auctions .........................................................37 3.2.2 How to alleviate the problems?..........................................................................42 3.2.3 Designing Combinatorial Auction Mechanisms ..............................................51 3.2.4 Combinatorial Auctions in Practice ..................................................................66
4 MULTI-ATTRIBUTE AUCTIONS................................................................69
4.1 The Scoring Function Approach................................................................70
4.2 Other Approaches .......................................................................................72
4.3 Multiple Attributes in Combinatorial Auctions .........................................75
5 ONLINE AUCTIONS AND AUCTIONS IN PRACTICE.................................77
5.1 Additional Design Issues .............................................................................79
5.2 Auction Rules ..............................................................................................81
III MOTIVATION AND OBJECTIVES ..............................................................86
ii
6 MOTIVATION AND OBJECTIVES ..............................................................86
6.1 Need for Mechanisms for Multiple-Unit Combinatorial Auctions ...........86
6.2 Identification of the Puzzle Problem and Need for Quantity Support......87
6.3 Objectives and Methods of the Study.........................................................89
IV DESIGNING AND TESTING THE QUANTITY SUPPORT MECHANISM....90
7 THE QUANTITY SUPPORT MECHANISM.................................................90
7.1 The Quantity Support Problem..................................................................91
7.2 An Example of a Combinatorial Auction with the QSM...........................94
8 TESTING THE QUANTITY SUPPORT MECHANISM: SIMULATION
STUDIES....................................................................................................98
8.1 First Simulation Study ................................................................................98
8.1.1 First Phase: Setting Up the Auctions.................................................................99 8.1.2 Second Phase: Simulations .............................................................................103 8.1.3 Third Phase: Benchmarks................................................................................104 8.1.4 Results of the First Simulation Study ..............................................................104
8.2 Second Simulation Study .........................................................................109
8.2.1 Phase One: Parameters ....................................................................................110 8.2.2 Phase Two: Simulations ..................................................................................114 8.2.3 Benchmark Cases ............................................................................................115 8.2.4 Results of the Second Simulation Study .........................................................117
9 IDEAS TO IMPROVE THE QSM ...............................................................132
9.1 Varying the Bid Decrement ......................................................................132
9.2 Initial Bids in the Form of Ranges ............................................................135
9.3 Allowing for Shortages and Excesses in the Supply..................................136
V THE GROUP SUPPORT MECHANISM.....................................................139
10 THE GROUP SUPPORT MECHANISM.....................................................139
10.1 The Group Support Problem....................................................................140
10.2 Customizing Cost Functions Approximations .........................................142
10.3 Example of an Auction with GSM ...........................................................145
10.3.1 Initiation Phase ................................................................................................146 10.3.2 The Auction.....................................................................................................149 10.3.3 Comparison with the Efficient Allocation ......................................................154 10.3.4 Comparison with the QSM Auction ...............................................................154
VI COMBIAUCTION – IMPLEMENTING THE QUANTITY SUPPORT MECHANISM IN PRACTICE..................................................................156
11 THE COMBIAUCTION SYSTEM...............................................................157
11.1 Organizing an Auction in the CombiAuction System .............................157
11.2 Bidding in the CombiAuction System......................................................163
iii
12 TESTING THE COMBIAUCTION SYSTEM ..............................................170
12.1 Experiment Set-Up....................................................................................170
12.2 Results of the Auction Experiment and Observations ..............................175
12.2.1 Efficiency of the Final Allocation and Total Cost to Buyer............................175 12.2.2 Bidding Strategies ............................................................................................182 12.2.3 Other Observations on Bidders’ Behavior and the CombiAuction System....193
VII CONCLUSIONS.........................................................................................196
REFERENCES.................................................................................................203
APPENDIX 1: Fixed Cost Parameters in the First Simulation Study ..................213
APPENDIX 2: Formulation of the QSP with True Cost Functions.......................214
APPENDIX 3: Formulation of the Efficient Allocation Problem .........................218
APPENDIX 4: Bidders’ Cost Functions in the Laboratory Experiments ...............221
APPENDIX 5: Auction Experiment Participant Questionnaire ............................224
iv
LIST OF TABLES
Table 1 Summary of market settings ................................................................... 10
Table 2 Example of a multiple-unit Vickrey auction: bid prices for the bidders ...... 29
Table 3 Summary of combinatorial auction mechanisms....................................... 66
Table 4 Design of the first simulation study........................................................ 103
Table 5 Results of Experiment I: Performance indicators of the dual price quantity support (D) and random support (R) ........................................ 105
Table 6 Results of Experiment II: Performance indicators of the dual price quantity support (D) and random support (R) ........................................ 106
Table 7 Results of Experiment III: Performance indicators of the dual price quantity support (D) and random support (R) ........................................ 106
Table 8 Results of Experiment IV: Performance indicators of the dual price
quantity support (D) and random support (R) ........................................ 107
Table 9 Results of Experiment IVb: Performance indicators of the random
support (R) with parameters drawn from the range [0, 2000] ................ 107
Table 10 Mark-ups in Experiment I..................................................................... 108
Table 11 Mark-ups in Experiment II ................................................................... 108
Table 12 Mark-ups in Experiment III .................................................................. 108
Table 13 Mark-ups in Experiment IV .................................................................. 109
Table 14 Ranges for fixed cost parameters in “normal” economies of scope.......... 111
Table 15 Ranges for fixed cost parameters in “large” economies of scope............. 112
Table 16 Total cost to the buyer as ratio of efficient production cost..................... 119
Table 17 Efficiency ratios of final allocations in quantity and price support auctions ............................................................................................... 121
Table 18 The ratio of total cost to the buyer in a quantity support auction and the lowest total cost in a non-combinatorial auction..................................... 123
Table 19 Average improvement in efficiency in the quantity support auctions ........ 130
Table 20 Average improvement in efficiency (as % of efficient cost) broken down according to three design parameters .................................................... 131
Table 21 Average final efficiency ratios broken down according to three design parameters........................................................................................... 131
Table 22 The initial bids and provisional winners ................................................. 147
Table 23 Initial estimates for the cost function parameters................................... 148
Table 24 Bids suggested by the GSM................................................................... 151
Table 25 Bids suggested by the GSM in Round 2.................................................. 151
Table 26 Cost function parameters after Round 2 ................................................ 152
Table 27 Bid stream and solution of the WDP after Round 4................................ 153
v
Table 28 The efficient allocation of the example auction ...................................... 154
Table 29 The final allocation of the auction with QSM ......................................... 155
Table 30 Winning bidders and bids in the equal capacities (A) auctions................. 176
Table 31 Total cost to buyer and efficiency indicators from the A auctions ............ 177
Table 32 Winning bidders and bids in the unequal capacities (B) auctions ............. 179
Table 33 Total cost to buyer and efficiency indicators from the B auctions ............ 180
Table 34 Frequencies of bidder strategies in the experiment auctions .................... 189
Table 35 Upper and lower limits for the uniform distributions from which the fixed cost parameters were drawn in the first simulation study................ 213
Table 36 Bidders’ cost function parameters in the equal capacities (A) auctions .... 222
Table 37 Bidders' cost function parameters in the unequal capacities (B) auctions.223
Table 38 Bidders' capacities in the unequal capacities (B) auctions....................... 223
LIST OF FIGURES
Figure 1 Cumulative distributions of the efficiency (compared to the production cost of the efficient allocation) of auctions with equal capacities and unequal capacities .............................................................................. 128
Figure 2 Cumulative distributions of buyer's total cost (compared to the production cost of the efficient allocation) in auctions with equal capacities and unequal capacities ........................................................ 129
Figure 3 Screen view from Step 1 of the auction creation process ...................... 158
Figure 4 Screen view from Step 2 of the auction creation process ...................... 159
Figure 5 Screen view from Step 3 of the auction creation process ...................... 160
Figure 6 Screen view from Step 4 of the auction creation process ...................... 161
Figure 7 Screen view of the auction owner's auction home page ......................... 162
Figure 8 Screen view of the list of open and scheduled auctions. ......................... 164
Figure 9 Screen view of the auction information page........................................ 164
Figure 10 Screen view of the bidder's "My Auctions" site...................................... 165
Figure 11 Screen view of the bidder's auction home page ..................................... 166
Figure 12 Screen view from the support tool dialog page ..................................... 167
Figure 13 Screen view of the Price Support tool.................................................. 168
Figure 14 Screen view of the bid suggestion made by the QSM............................. 169
Figure 15 The organization of the laboratory experiment..................................... 175
1
I INTRODUCTION
Almost everyone is familiar with the concept of an auction. When thinking of an
auction, we immediately imagine a room full of people at an auction house such as
Sotheby’s or Christie’s. The broker presents the item for sale – typically an expensive
work of art – and announces the starting price. Bidders with numbered signs in their
hands indicate their willingness to purchase the item as price rises. Today, however,
the auction is not limited in time or space. The advent of the Internet allowed the
transferring of the auction house online. Online auctions have become increasingly
popular, and today eBay is at least as well known as Sotheby’s or Christie’s. The variety
of products up for auction is incredible: you can buy pretty much anything from
dinosaur eggs to real estate property, and from diamonds to used clothes. An increasing
number of people have also participated in an auction themselves, either traditional or
online.
The concept of an auction, however, is not as simple or narrowly defined as a typical
bidder in an eBay auction may think. Imagine you are a philatelist, and that you are
attending a stamp auction in hopes of adding to your collection of 19th century Finnish
stamps. You are well aware of the fact that the stamps are more valuable as a complete
series than individually. Thus, you would be willing to pay a lot more for the two
stamps missing from one of your series than for them individually. Now, if in the
auction all stamps were auctioned individually, one after the other, how much would
you be willing to bid for the first one when you do not know whether you will win the
second one? Would you not be happier, if you could indicate to the seller that you
would be willing to pay more, if you were guaranteed both stamps? The question then
becomes, why would the seller be selling the stamps individually and not as a series?
Because not all bidders are interested in the whole series, but rather different subsets of
the series, and the seller cannot know what kind of packages to build from the stamps.
This simple example demonstrates that there are several cases in which the traditional
Sotheby’s style single-item auction is not optimal for the seller or the bidders. One is
then tempted to ask, if the auction could somehow be modified to deal with the
2
problem in this example. Fortunately, combinatorial auctions provide a solution to the
problem, as we will see later in section 3.2.
The typical single-item forward auction prevalent in eBay, Sotheby’s and Christies is
only one of many different classes of auctions. McAfee and McMillan (1987) define an
auction as “a market institution with an explicit set of rules determining resource
allocation and prices on the basis of bids from the market participants.” This definition
is quite generic, and later on in this thesis we will see, how different kinds of
mechanisms fall under the category of auctions. Auctions are often thought of as a
special case of negotiations, with more strict structure and a set of rules than one-on-
one negotiations. Some researchers even use the term “negotiation” to also include
auctions, but in this thesis, the term negotiation will refer only to one-on-one
negotiations. With auctions I refer to bidding processes between one seller and many
buyers or one buyer and many sellers. Specifically, the case with one seller and many
buyers is called a forward auction, and the case with one buyer and many sellers is
called a reverse auction or a tender. Forward auctions are common in business-to-
consumer (B2C) or consumer-to-consumer (C2C) transactions, whereas reverse
auctions are often used in business-to-business (B2B) transactions. When there are
multiple sellers and multiple buyers in the market, it is called a double auction (in
which both sellers and buyers place bids, like in the stock market), or it is simply a
regular market characterized by fixed prices.
The purpose of auctions and negotiations is to determine a price for the item(s) in
question. They allow the seller to discover the buyers’ valuation (and the correct price)
during an auction process. Thus, auctions are good for selling and buying non-standard
products, the market price of which is difficult to know beforehand.
Even though most people are familiar with the concept of an auction, the field of
auction studies – the science of auctions – however, is not as well or broadly known.
The question “Why study auctions?” is legitimate, and should be answered to justify
any research in the area. McAfee and McMillan (1987) argue that the fact that they
exist in practice is reason enough to study them. A rigorous theoretical treatment of
auctions can help design better auctions in practice. Moreover, studying people’s
3
behavior in auctions empirically can help design better auctions. McAfee and
McMillan (1987) also argue that auctions are a case of price setting under imperfect
and possibly asymmetric information that is interesting in a game theoretic sense.
Maasland and Onderstal (2006) mention that auctions are worth studying, because
they can be applied to other situations as well, e.g. modeling incentive contracts.
Auction-based approaches have also been used in scheduling and the design of
manufacturing systems (Srivinas et al., 2004, and Kumar et al., 2006), and in
distribution of computing resources over a grid of interconnected computers
(Schnizler et al., 2008). Thus, advances in auction theory and design can have broad
implications.
Systematic and scientific study of auctions began surprisingly late, considering how
long auctions have been used in transactions. Auction research started essentially in
1961 with William Vickrey’s seminal paper. At the beginning, auctions were of interest
to only a small group of economists, who saw them as interesting applications of game
theory. Since then, the field of auction studies has grown in both depth and breadth to
the wide multidisciplinary field it is today. Economics and game theory still have a
strong foothold in auction research, but there is a growing interest among computer
scientists, mathematicians, and decision analysts towards auctions.
As the field of auction studies grew, also the concept of auctions was broadened. At first
only traditional single-item, single-unit, price-only auctions were considered (see
section 1.2 for definitions). Then researchers started to consider more complicated
designs: multi-unit auctions, multi-attribute auctions, and finally combinatorial
auctions (which are a case of multi-item auctions). The new designs allowed more and
more goods to be sold and bought through auctions. The Internet has enabled the
implementation of many of the more complicated designs, and it has brought up new
research questions. Today, over forty years after Vickrey’s article, both the use of
auctions as a market mechanism, and the study of auctions are thriving.
This thesis contributes to the study of combinatorial auctions. In short, combinatorial
auctions are multiple-item auctions, in which the bidders can place bids on packages
of items. I discuss the complexities of combinatorial auctions, and the need for
4
decision support for bidders in such auctions. In particular, the contribution of this
thesis is the design of decision support tools for bidders in semi-sealed-bid, iterative
combinatorial auctions (see section 6.3 for a more elaborate description of the
objectives and methods of this study). The decision support tools are algorithms that
provide the bidders with suggestions for bids to be placed at any given time in the
auction. In this thesis I present the mathematical formulation for the decision support
tools, the Quantity Support Mechanism (QSM), and its variation, the Group Support
Mechanism (GSM). These tools are also easily extended to a multi-attribute
combinatorial auction. I also present the design and results of two simulation studies
we ran to test the QSM. Based on the results of the studies, some improvement ideas
were developed, among them the GSM. I will present the improvement ideas, and
mathematical formulations for them. The QSM was also implemented in an online
auction system, CombiAuction. The user interface and the usability of the QSM were
tested in a laboratory experiment with human subjects. Thus, this thesis also
contributes to the design of auctions in practice. Also, I identified strategies bidders
used in the combinatorial auctions in the experiment.
The structure of this thesis is the following. The Introduction contains this general
introduction to auctions and auction research to be followed by a chapter reviewing the
key concepts related to auction design (market types, auction mechanisms). The
second part reviews relevant literature on auctions. The purpose of the literature review
is to discuss combinatorial auction research, to position it in the field of auction
studies, and to illustrate the links between other disciplines (economics, computer
science, multicriteria decision making) and auction research. The third part discusses
the need for support in combinatorial auctions, and shows the gaps in existing
literature that this study attempts to fill. The third part also contains a more detailed
description of the objectives and methods of the research. The fourth part presents the
formulation of the QSM, and the designs and results of two simulation studies. The
chapter discussing possible small improvements to the QSM concludes the fourth part,
and demonstrates the need for more drastic changes in the QSM if significant
improvement is to be achieved. The fifth part introduces the Group Support
Mechanism (GSM), a support tool based on the QSM but with a few significant
5
alterations, and presents an example to illustrate the mechanism. The sixth part
describes the CombiAuction, an online auction system in which the QSM has been
implemented, and an experiment I ran with human subjects. The seventh part
concludes the thesis.
6
1 DESIGNING AUCTIONS
Understood as a traditional arts auction, auctions seem to be relatively simple market
mechanisms. However, the auctions are actually a group of different transaction
mechanisms, some more complicated than others. The common feature with all the
transaction mechanisms categorized as auctions is that their purpose is to determine
the price and allocation of the item(s) in question. Researchers have identified a
number of desirable properties for auctions, and the purpose of auction design is to
ensure that a desirable outcome is reached. Auction design takes place on many levels.
On the macro level, the market setting, i.e. the type and number of goods for sale,
affects the design. In addition, the auction owner must decide on the appropriate
auction mechanism (e.g. ascending price or descending price, open or sealed-bid).
Under any market setting, a number of different auction mechanisms can be used. On
the micro level, auction design must also say something about the detailed design
aspects (whether to set a reservation price, what kind of information to disclose, etc.)
and the rules of the auction (who can participate, when will the auction end, how is
defaulting by a bidder dealt with, etc.) to ensure fairness, maximal revenue for the
auction owner, and impede cheating by the participants. In this chapter I will first
review the desirable properties of auction designs. Thereafter I will review macro level
concepts and definitions related to auction design including different market settings
and auction types, and the most common auction mechanisms. The details of auction
design and auction rules will be discussed more thoroughly in the context of Internet
auctions (Chapter 5 of the literature review).
1.1 Desirable Properties of Auctions
There are several properties the auction organizer may want the auction to possess.
From the economics point of view, an important property is allocative efficiency (or
efficiency for short). Allocative efficiency is reached when the winner in the forward
auction is the bidder with the highest valuation (or lowest production cost in reverse
auctions). This also means that total welfare in the society is maximized. Several
researchers (e.g. Bichler, 2001, Milgrom and Weber, 1982, and McAfee and
McMillan, 1987) use the term “Pareto optimality” in a similar sense as allocative
7
efficiency, but in my opinion it is not the best term to use here, and also Koppius and
van Heck (2002) make a distinction between Pareto optimality and allocative
efficiency. Thus, in this thesis I will use the term allocative efficiency in the context of
auction design.
A property closely related to allocative efficiency is revenue maximization (or cost
minimization). Often the seller would like to design the auction in such a way that her
revenue is maximized. An auction mechanism that maximizes the seller’s revenue is
also called an optimal mechanism. Usually a revenue maximizing auction is also
efficient, because the bidder with the highest valuation is the one who is willing to pay
the most. However, an efficient auction need not be revenue maximizing, as we shall
see later on in section 2.5.
The objective of incentive compatibility is also closely related to allocative efficiency
and revenue maximization. Incentive compatibility means that the bidders have no
incentive to shade their bids, and that they would be willing to report their true
valuations, because lying would not increase their pay-off. When an auction is
incentive compatible the bids reflect the magnitude of the bidders’ valuations, and the
ranking of the bids reflects the ranking of the bidders’ valuations. Thus, the winner is
the one with the highest valuation, and depending on the payment rule, the revenue to
the bid taker can be maximized.
Other desirable properties include fairness, failure-freeness (robustness), resistance to
cheating and manipulation, and low transaction costs. These properties are discussed
in more detail in the literature review (Chapter 5).
1.2 Market Settings
The market setting under consideration sets the stage for the auction design. The
different market settings can be classified based on 1) the number of different items
(products or services) to be sold1 in one auction, 2) the number of homogenous units of
each item, and 3) the number of units each bidder wishes to acquire. In the simplest
1 I will present the auction concepts in the context of forward auctions, which is the convention in auction literature. All the concepts apply in the reverse setting, with the only difference that items in the auction are to be bought (not sold), and thus bids represent supply rather than demand.
8
case there is only a single unit of one item for sale. Naturally, in this setting the bidders
can only demand one unit. This auction type is often called a single-item auction for
short. Single-item auctions were the focus of most early auction studies conducted by
economists and game theorists. The properties of single-item auctions will be discussed
in Chapter 2 of the literature review.
A logical extension of the single-item auction is the (single-item) multiple-unit
auction. In such an auction at least two identical units of a particular item are to be
sold. Depending on the nature of the item, each bidder can now demand either one
unit (for example, if several identical licenses are auctioned, the bidders would want at
most one license), or several units. See section 2.8 in the literature review for more
discussion on multiple-unit auctions.
The most complex auctions are the multiple-item auctions, which are the focus of this
thesis. There can be either one unit or multiple units of each item for sale. The bidders
may wish to acquire only one of the items, or several. However, it is common to
assume that at least one of the bidders wish to acquire more than one item in the
auction. If each bidder had a demand for only one item, it is difficult to imagine, why
hold one auction for a collection of items, unless the items were close substitutes.
Thus, the case of multiple items but single-item demand is usually neglected in
literature. Theoretically, in multiple-unit, multiple-item auctions bidders can either
demand one unit or multiple units of the different items. Usually, however, in
multiple-item auctions it is realistic to assume that bidders demand more than one
unit. Thus, the case of multiple items and multiple units, but only single-unit demand
has not been considered in auction literature.
In the multiple-item auctions there is also a big difference in the design of the auction
depending on how the bidders are allowed to express their demand. Basically the
choice is between allowing bids on combinations of items (also called package
bidding and combinatorial bidding), or not allowing them. A combinatorial bid is a
vector containing the desired quantities of each item, and a single price for the
combination. The efficiency of the auction is improved, if bids on combinations are
allowed, but the complexity of the auction increases a lot. I will return to combinatorial
9
auctions and this trade-off in section 3.2. The different market characteristics described
above result in eight different market settings (see Table 1).
Thus far I have implicitly assumed that bids are evaluated only based on the price
attached to the bid. However, just as in many negotiations, the price of the items is not
necessarily the only attribute of interest to the buyer in an auction. In such cases, the
ability to make multi-attribute2 bids can increase the efficiency of the auction, and it
can even make it rational to sell through auctions some items, which have previously
been sold through one-on-one negotiations. In these auctions, the bidders’ bids are
multidimensional vectors with one component for each attribute. Examples of non-
price attributes often used in multi-attribute auctions are quality, terms of payment,
delivery times etc. Multi-attribute auctions can be either single-unit or multiple-unit
auctions. Guttman and Maes (1998) refer to multi-attribute auctions as win-win
situations since the auction is no longer a zero-sum game and it is possible for both
sellers and buyers to be better off. Thus, the nature of the multi-attribute auction is
somewhat different compared to the price-only auctions. The addition of multiple
attributes to the auction causes new kinds of complications in the auction process. For
instance, the comparison of bids against each other becomes conceptually difficult.
The characteristics of multi-attribute auctions, the arising problems, and the attempts
to overcome the problems are briefly discussed in Chapter 4.
The possibility to consider other attributes besides price adds another eight different
variants to the original eight market settings. Thus, every row in Table 1 corresponds to
two market settings: one with price as the only bid attribute, and the other identical
otherwise but with multiple attributes considered.
2 Some researchers use the term ‘multi-dimensional auction’, but I will use the term ‘multi-attribute auction’ to avoid confusion, because former term can also refer to a combinatorial auction.
10
Table 1 Summary of market settings
# of Items # of Units Demand Package Bidding # of Attributes
single single single-unit - price-only/ multiple attributes
single multiple single-unit - price-only/ multiple attributes
single multiple multiple-unit - price-only/ multiple attributes
multiple single single-item not allowed price-only/ multiple attributes
multiple single multiple-item not allowed price-only/ multiple attributes
multiple multiple multiple-unit, multiple-item
not allowed price-only/ multiple attributes
multiple single multiple-item allowed price-only/ multiple attributes
multiple multiple multiple-unit, multiple-item
allowed price-only/ multiple attributes
1.3 Auction Mechanisms
An auction mechanism is defined as a set of rules telling how the winner is
determined, how the payments of each bidder are determined, and how the bid
information is collected from the bidder. According to Krishna (2002), a generic
mechanism consists of three elements: the set of possible bids, the allocation rule, and
the payment rule. The allocation rule determines the probability with which a bidder
will win the object. The payment rule determines the payment the bidder with the
winning bid must make. In every market setting described above there can be different
auction mechanisms. Auction literature recognizes four basic mechanisms, which are
most commonly used: the English auction, the Dutch auction, the first-price sealed-
bid auction, and the second-price sealed-bid auction (also known as the Vickrey
auction after its inventor William Vickrey). These basic mechanisms are special cases
of the generic mechanism.
11
In a traditional ascending price English auction, the auctioneer3 starts the bidding at
the reservation price, if the auctioneer has set one. If no reservation price is specified,
the starting price is set to the starting price specified by the seller. If no such price is
determined, the starting price is zero. The bidders can then call out bids. A new bid
has to exceed the currently highest bid to be acceptable. Depending on the specific
rules of the auction, the bidders can either freely call out any acceptable bid, or the
auctioneer calls out new prices which the bidders can accept. The latter version is
referred to as a clock auction. The auction ends when no bidder is willing to increase
her bid. The bidder i with the highest bid wins the item and pays the price equivalent
to her bid. English auctions are most common in practice. They are especially popular
in art and antiquities auctions, and consumer-to-consumer (C2C) auctions, such as
eBay.
The Dutch auction reverses the logic of the English auction: the price descends in the
Dutch auction. Thus, in the Dutch auction the auction clock is set at a very high price
at the beginning. In fact, the price is set so high that no bidder would be willing to pay
the price. When the auction begins, the price indicated by the auction clock is
gradually decreased until one bidder indicates her willingness to pay the current price.
The auction ends, and the bidder receives the item at the price indicated by the
auction clock. If the auction clock reaches the seller’s reservation price, and no bids
have been made, the item is left unsold. The allocation rule in the Dutch auction is
the same as is the English auction: the bidder with the highest bid wins the item.
However, in the Dutch auction the allocation rule is trivial, since by definition there
will only be one bid in the auction. Dutch auctions have been used a long time in
flower auctions in the Netherlands (hence the name). Also other perishable items (e.g.
fish) are auctioned through Dutch auctions, because they are fast to conduct.
In the first-price sealed-bid auction4, all bidders simultaneously submit their bids to the
auctioneer. This means that the bidders are unaware of the content of all bids except
3 In this text the “auctioneer” or the “auction owner” refers to the bid taker, i.e. the seller of the good in a forward auction and the buyer in a reverse auction. It is also possible that the auctioneer is a neutral third party (e.g. an auction house) but that case is omitted from this discussion. 4 In this text I will use “first-price auction” as shorthand for “first-price sealed-bid auction.”
12
their own. The auctioneer goes through the bids and the bidder with the highest bid
wins the item and pays the price equal to her bid (if the price exceeded the seller’s
reservation price). The allocation and payment rules are the same as in the Dutch
auction. The second-price sealed-bid auction is identical to the first-price auction
except for the fact that the winner (the bidder with the highest bid) pays the amount
equal to the second highest bid. Sealed-bid auctions are a common practice in
procurement situations, both public and private sector. One reason for their popularity
is the fact that the bids may contain critical information of the bidders’ cost structure or
their competitive advantage. Thus, bidders prefer to keep their bids secret from their
competitors.
The English auction and the Dutch auction are open-cry auctions as opposed to sealed-
bid auctions. In open-cry auctions the bidders call out their bids. In other words,
bidders are aware of the actions of their competitors. Dynamic (progressive, iterative)
auctions have multiple rounds and bidders can revise their bids (place several bids).
The English auction is a dynamic auction. The Dutch, first-price and Vickrey
auctions, on the other hand, are static auctions. In all three auctions the bidders have
only one chance to place the bid and no revision is allowed.
All the four mechanisms presented above were considered in the forward setting.
However, all mechanisms can be used in the reverse setting as well. If in the forward
English auction the price was ascending, it is descending in the reverse English
auction. Similarly, in the reverse sealed-bid auctions the winner is the bidder with the
lowest bid. In the reverse Dutch auction, the clock starts at a very low price and is then
increased until some bidder agrees to take the item (e.g. a contract) at the current
price. In the reverse first-price and Vickrey auctions the winner is the bidder with the
lowest bid, and the price in the Vickrey auction is that of the second lowest bid.
Even though I have presented the four basic auction mechanisms only in the simple
single-item, single-unit case, the same mechanisms can be extended to more complex
settings. Also, most auction mechanisms considered in literature or used in practice
contain elements of one or more of the four basic mechanisms, as will become
apparent in the literature review.
13
II LITERATURE
2 THE ORIGINS OF AUCTION THEORY: SINGLE-ITEM
AUCTIONS
Traditional auction theory discusses mainly the single-item, single-unit auction. The
setting is simple enough so that equilibrium strategies can be solved analytically and
comparisons between different auction mechanisms can be made. A good reference for
a pure game theoretic discussion of auctions is Wilson (1992). However, the game
theoretic approach requires that some specific assumptions have to be made about the
bidders and the item on sale. Literature focuses on two main models that differ
significantly in their assumptions about the bidders’ valuation for the auctioned item:
the Independent Private Values (IPV) model and the Common Value (CV) model. In
the IPV model each bidder’s valuation is assumed independent of other bidders’
valuations. The bidder, however, knows her valuation with certainty. In the CV model
the item’s value is the same for all bidders, but the bidders only have an estimate of the
item’s “true” value. The affiliated values model presented in section 2.3 combines
elements from the two extreme models. Studies are concerned with the expected
revenue from the four basic mechanisms, the effects of relaxing assumptions behind
the model, and the design of optimal auction mechanisms (i.e. revenue maximizing
mechanisms) given the set of assumptions in the models. Both IPV model and the CV
model, as well as the affiliated values model will be presented briefly in the following
sections.
A logical extension of the single-item, single-unit auction is the single-item, multiple-
unit auction. I will call it the multiple-unit auction for short. In such auctions, a set of
identical objects is sold (bought). The models for multiple-unit auctions are usually
created under the IPV assumptions. Multiple-unit auctions are discussed in section 2.8.
All the auctions in this section are presented in the forward setting, because it is
common practice among economists. I will try to keep the presentation non-technical,
so the reader interested in the exact proofs of propositions is advised to look them up in
the original references.
14
2.1 The Independent Private Values Model
The Independent Private Values (IPV) model was the first framework in which
auctions were systematically studied. The earliest studies in the spirit of the IPV model
were conducted already in the 1950s. Friedman (1956) analyses a situation in which
each bidder can estimate the other bidders’ behavior based on prior experience. The
groundbreaking work was done by William Vickrey (1961). His main discovery was
that under specific assumptions the second price auction he designed would generate
the same expected revenues as the first-price and Dutch auctions.
The key assumptions behind the IPV models are (collected from Vickrey, 1961,
McAfee and McMillan, 1987, Rothkopf and Harstad, 1994):
1) Each bidder knows the true value of the item for her, but she does not know the valuations of the other bidders.
2) The valuation of one bidder is statistically independent of any other bidder’s valuation.
3) The bidder perceives the other bidders’ valuations as drawn from some known probability distribution, and she knows that other bidders regard her valuation as being drawn from some distribution.
In addition, some other assumptions are either explicitly or implicitly presented in the
context of the IPV model (see e.g. McAfee and McMillan, 1987, Rothkopf and
Harstad, 1994, Maasland and Onderstal, 2006):
4) The bidders (and the seller) are risk-neutral. 5) The bidders are symmetric (i.e. they draw their valuations form the same
distribution). 6) There is a single, isolated auction (not a multiple-stage game), and the number
of bidders participating is fixed. 7) There is no collusion among bidders. 8) There are no externalities from the allocation of the item to the bidder, or from
the payment made by the winner.
2.1.1 Strategic Equivalence of Auction Mechanisms in the IPV Model
In the IPV model, the Dutch auction and the first-price auction are strategically
equivalent, as are the English and second-price auctions. The strategic equivalence of
the Dutch and first-price auctions actually extends to other situations as well, but the
strategic equivalence of the English and second-price auctions breaks down if bidders’
15
valuations are not independent (see sections 2.2 and 2.3 for more discussion).
However, for now let us return to the IPV model.
Several review papers provide a thorough explanation of the strategic equivalence. The
following presentation is adapted from Paul Milgrom (1989), Martin Bichler (2001),
and the somewhat more mathematical presentation of McAfee and McMillan (1987).
Consider first the English auction. The auction starts at a relatively low price and the
price rises as the auction proceeds. Since the winning bidder has to pay her own bid,
and every bidder knows her own valuation, it does not make sense for them to bid a
price above their valuation (they would be better off not participating at all than paying
more than their valuation). One by one, bidders drop out of the competition until only
two bidders are left. I will denote them with B1 (bidder with the highest valuation) and
B2 (bidder with the second highest valuation). Once the current bid price reaches the
valuation of B2 she will not place any more bids. Bidder B1 could still bid higher, but it
does not make sense to do so since she can win the auction by bidding only marginally
higher than what B2 is willing to bid. Thus, if B1 bids rationally (and we will assume
that she does), she wins the item and ends up paying the price equal (or almost equal)
to the valuation of B2. The dominant strategy for each bidder is to bid until the price
reaches her valuation – regardless of what other bidders do.
The Vickrey auction seems different to the bidders since they cannot observe each
others’ bids and therefore obtain no information on other bidders’ valuations. However,
it does not matter, because as was explained in the previous paragraph, the opportunity
to observe competitors’ behavior in the English auction did not affect the bidders’
optimal strategy. As it happens, in the Vickrey auction it is also a dominant strategy for
each bidder to bid their own valuation. Remember that the bidder who wins only has
to pay the price indicated by the second highest bid, i.e. the price is determined
independent of the winning bid. If the bidder bids less than her valuation, she gains
nothing with that move: if she is still the winner, she pays the price of the second
highest bid, which would have been the same even if she had bid her valuation. Thus,
by bidding less than her valuation the bidder only risks losing the item. If, on the other
hand, the bidder bids higher than her valuation, she increases her chances of winning
16
the auction. This strategy becomes effective only when she is not the bidder with the
highest valuation, and by bidding higher than her valuation she manages to outbid the
bidder B1. She manages to win the auction, but the price she has to pay (i.e. the second
highest bid) is now equivalent to the valuation of bidder B1. This price, by definition, is
higher than the winning bidder’s valuation of the item, and she ends up with a negative
pay-off. Thus it is clear that rational bidders bid a price equivalent to their valuation.
The bidder with the highest valuation wins, and she pays the price equivalent to the
valuation of the bidder with the second highest valuation. The strategies and the
outcome are clearly the same as in the English auction.
The Dutch and first-price, sealed-bid auctions seem very different from one another at
first glance. In the former, the auctioneer cries out prices whereas in the latter each
bidder submits a sealed bid. However, assuming that the bidders plan their actions
prior to the auction, their decision problem is exactly the same in either case. Let us
assume that the bidder is bidding for an item in a Dutch auction. Assume that her
strategy is the following. First she waits for the price to drop to p1 and if no one has
claimed the item, she will either place a bid, or wait. Assume that p1 is greater than
what she is willing to pay for the item, so she decides to wait. She now chooses p2 as her
new point of evaluation, and the same process repeats itself. As long as the price is
higher than her valuation for the item, the choice is trivial. Once the price goes below
her valuation, she needs to weigh the added utility from letting the price fall further
against the risk of losing the item to another bidder. Finally, at price p the bidder
places a bid and claims the item. Notice, however, that she has made her choices
always under the assumption that no one else claimed the item. There is no point in
considering the case when some one places a bid before her, because then the game is
over. In the beginning of the auction she could have been asked to directly indicate the
highest price at which she is willing to claim the item, and the answer would have
been the same. The only additional information the bidder gets during the auction is
either that somebody was willing to pay more, or that other bidders were not willing to
pay as much as she was. Either type of information is useless to a bidder, when
valuations are assumed independent and private. In the first-price auction the bidder is
faced with the same trade off between larger utility and smaller probability of winning,
17
because the Dutch auction offers no added information to the bidders in the IPV case.
Thus she ends up bidding the same price p for the same item. This can be verified
mathematically.
I have now shown that the price paid by the winner in the English and second-price
auctions is equal to the second highest valuation, and in the Dutch and first-price,
sealed-bid auction equal to the expected value of the second highest valuation. The
auction owner organizing the auction does not know the valuations of the bidders, so
her expected revenue from all four auction mechanisms is always equal to the expected
value of the second highest valuation! Note that this does not imply that the actual
outcomes of all auction mechanisms would always be the same. In the English and
Vickrey auctions the price paid by the winner is always equal to the valuation of bidder
B2. In a first-price and a Dutch auction the price of the winning bid is the expected
value of the second highest valuation. These two prices are identical only by accident,
but on the average they are the same.
2.1.2 The Revenue Equivalence Theorem
Vickrey’s result in 1961 regarding the revenue equivalence of the four basic auction
mechanisms (as described in the previous section) was the preliminary version of the
much celebrated revenue equivalence theorem. The revenue equivalence, however,
can be extended to a much broader class of auction mechanisms. The exact
formulation of the theorem was proposed by Myerson (1981). According to Myerson,
the seller’s revenue from the auction is completely determined by the allocation rule,
and the utility gained by the bidder with the lowest possible valuation. As long as the
auction mechanisms allocate the item to the same bidder, and the utility for a bidder
with the lowest possible valuation is the same, the expected revenue for the seller is also
the same.
Vickrey’s discovery of the equivalence of the four basic mechanisms is thus a special
case of the revenue equivalence theorem. In all four auctions the allocation
mechanism is the same (highest bid wins), and the expected utility of the bidder with
the lowest possible valuation is zero (she would never be the winner, and losing bidders
do not have to pay anything). Only the payment rules are different, but according to
18
Myerson’s theorem, the payment rule does not matter. One should keep in mind,
though, that Myerson’s theorem relies on the assumptions of independent private
values, bidder symmetry and risk neutrality. When the IPV assumptions are relaxed,
the equivalence of the four basic mechanisms breaks down. The practical implication
of this is that in reality it does matter, which mechanism the seller chooses.
2.2 The Common Value Model
Not satisfied with the assumption of independent and private values (assumptions 1
and 2), researchers such as Wilson (1969) and Capen, Clapp and Campbell (1971)
replaced it with an opposite assumption. The underlying assumption in the Common
Value model is that the true value of the item is the same to all bidders, but at the time
of the auction this value is unknown to all participants. This assumption is valid, for
example, in the bidding for oil drilling rights, where the value of the asset (the value of
the oil extracted) is the same to all participants, but no one knows beforehand the
amount of oil in the area. The bidders make estimates of the “true” value of the item
based on the information they have.
The interesting characteristic of a common value auction is that it is by definition
always efficient. The downside of a common value auction is that the task of the
bidders is much more difficult than in private value auctions. It is hard to determine a
bid, when you do not know the value of the item to you. If the bidders are not careful
they can easily fall prey to what is called the winner’s curse. The following example
presented by Milgrom (1989) illustrates the pitfall embedded in the common value
auction. The example is presented in the reverse setting contrary to the previous
sections. Let’s assume that contractors are invited to bid on a job, which each of them
can complete at a cost C. The contractors make unbiased estimates ci = C + εi of the
cost, where εi is the estimation error of contractor i. The estimation errors are
statistically independent, and the expected value of εi is zero (because estimates were
assumed unbiased). Even though on average the estimates are correct, in every auction
there will always be bidders whose estimates are higher or lower than the true value C.
Because the expected value of the error term is zero, the expected value of the smallest
estimate error must be less than zero. Each bidder is unaware of the other bidders’
19
estimates, so she cannot know whether her estimate is above or below average. The
bidder with the lowest cost estimate will bid for the lowest price and will win the
contract at that price. However, as she was the bidder with the smallest ci, her bid price
is likely to be less than the true cost of the contract C. The winning bidder incurs losses
and thus suffers from the winner’s curse. In other words, the bidder’s estimate of the
production cost increases when she learns that she is the winner (and therefore the one
with the lowest estimate). Unfortunately, at that point the auction has closed and she
cannot revise her bid.
A wise bidder acknowledges the fact that if she wins, she has underestimated the
production costs. Thus, she bids a higher price than bidders who do not acknowledge
it. The optimal bidding strategy is based on the assumption that the bidder is the one
with the lowest estimate, and the task is to, given this assumption, figure out the
expected value of the second lowest estimate. Equilibrium bidding strategies can be
derived just as for the IPV model, but they are a lot more complex (see, for example,
Wilson, 1977).
In the common value case the English auction differs from the sealed-bid auctions,
because the information available during the auction is now valuable to the bidders
(contrary to the IPV case). The bidders can observe others’ value estimates, and revise
their own estimates accordingly. Thus there is not as much reason to correct the bids as
the estimate becomes more accurate (Milgrom and Weber, 1982). Due to less
uncertainty, bidders can also bid more aggressively and the English auction leads to
higher expected revenues for the auction owner.
2.3 The Affiliated Values Model
In most cases it is not realistic to assume the bidders’ valuations to be strictly common
or independent. Items for sale have both a private and a common value element
(Bichler, 2001). Unique pieces of art are usually used as examples in the IPV models,
but more often than not the bidders are also interested in the resale value of the art
work in addition to their private valuation. Thus there is a common value element in
most IPV cases. Similarly, one can argue that firms differ in their resources and
capabilities, and thus there can be a private value element in common value auctions.
20
Milgrom and Weber (1982) introduce a more general model – the Affiliated Values
model – that combines elements from both the IPV and Common Value models.
Affiliation between valuations means that if some bidders value the asset highly, it is
more likely that the other bidders’ valuations are high as well. The exact mathematical
definition of affiliation is adopted from Milgrom and Weber (1982). Let v represent the
vector of the valuations of the N bidders of the auction. Let f: RN→R denote the joint
probability distribution of the valuations. Finally, let 'vv ∨ denote the vector
containing the component-wise maximum, and 'vv ∧ the component-wise minimum
of two valuation vectors v and 'v . Then
)'()()'()'( vfvfvvfvvf ≥∧∨ (1)
This equation is roughly saying that it is more likely to have either a vector with
relatively large valuations for all bidders ( 'vv ∨ ) or relatively low valuations ( 'vv ∧ )
than a mixture of high and low values (either v or 'v ).
The Affiliated Values model allows for statistical dependence between the bidders’
value estimates as well as for differences in individual tastes. There can also be different
degrees of affiliation. The IPV and the common value cases are included in the model
as two extreme cases alongside numerous intermediate models, which are perhaps
more realistic.
The revenue equivalence of the four basic mechanisms does not hold under the
affiliated values model. In fact, Milgrom and Weber (1982) show that the four different
mechanisms can be rank-ordered based on the expected revenue collected in the
auction. It can be shown that the English auction generates more revenue than the
other auction mechanisms. The situation is analogous to the pure common value
auction: in an open-cry auction the bidders obtain information about each other’s
valuations and can then revise their own estimates. It can also be shown that the
Vickrey auction yields a higher expected revenue than the Dutch and first-price
auctions, which remain revenue equivalent. Although the revenue ranking of affiliated
values auctions is similar to the pure common value auction, the observation that a
common value auction is always efficient does not carry over to the affiliated values
case (Goeree and Offerman, 2002). It is interesting that also the IPV auctions are
21
theoretically efficient, yet the affiliated values auction, which is a mixture of both
extremes, does not have this property. The intuition behind this result is quite
straightforward. When both private and common elements are present, a bidder with a
small private value but an overly optimistic estimate of the common value element
may outbid a bidder with a higher private value.
2.4 The Popular English Auction
In light of the perhaps surprising results of the revenue equivalence theorem, it is
interesting to observe that in practice auction owners are not indifferent between the
auction mechanisms. For instance, the English auction is by far the most popular, and
the Vickrey auction is hardly ever used. Milgrom (1989) offers an explanation to this
discrepancy by saying that revenue is only one criterion to evaluate auction
mechanisms. Others are robustness, efficiency, transaction costs, fairness, and
immunity to cheating.
Robustness here refers to the vulnerability of the mechanism to changes in the IPV
model assumptions. The English auction and the Vickrey auctions are more robust
than the Dutch and first-price auctions. This is because bidders have a dominant
strategy, which is independent of the distribution of other bidders’ valuations, and the
number of bidders in the auction (McAfee and McMillan, 1987). On top of that, the
English auction is better than the Vickrey auction when there is a common value
element in the auction. This is because the English auction is the only mechanism of
the four basic mechanisms in which the bidders can observe each other’s bids, and
learn about other bidders’ valuations. However, introducing risk averse bidders in an
auction makes the first-price and Dutch auctions better in revenue terms (McAfee and
McMillan, 1987). And since risk aversion among bidders is not an unreasonable
assumption, this is a valid argument against the English auction. Another such
argument is that the English auction is more vulnerable to collusion in the form of
bidding rings (Robinson, 1985), and also cheating in the form of signaling simply
because it is the only mechanism in which the other bidders can see the content of the
bids.
22
Efficiency usually refers to the allocative efficiency of the auction outcome, which is
always achieved under the IPV model. However, relaxing some of the assumptions
breaks this result for the Dutch and first-price auctions. Milgrom (1989) broadens the
concept of efficiency to include bid preparation costs. The more information gathering
is required, and the more complicated the calculation of the bidding strategies, the
higher the preparation costs. This argument also favors the English and second-price
auctions, as the bidding strategies are simple, and in English auctions all the
information available can be collected during the auction. According to Engelbrecht-
Wiggans (2001) the lower participation cost of English auctions attracts more bidders,
and on average the more there are bidders, the higher the expected revenue for the bid
taker.
On the other hand, the preparation cost of bids in the English auction may be low, but
it requires the bidders to actively participate in the auction for the whole duration of
the bidding process. In 1989, when Milgrom wrote his article, participation also usually
required physical presence. Nowadays with the Internet, physical presence is not that
critical, but the English auction still requires the bidder to be alert and present online.
Thus there is a time cost involved in the English auction, which increases its otherwise
low transaction costs. A quick remedy for this would be to use bidding agents. The
bidder would indicate the highest acceptable price, and the agent would bid on her
behalf up to that price. In fact, bidding agents of one sort or another are becoming
more and more common. For example, eBay uses bidding agents. Another solution
would be to organize a Vickrey auction – this is the only solution, if we are talking of a
traditional “off-line” auction. This is not usually done, however, because the Vickrey
auction is highly susceptible to manipulation. Nothing stops the bid taker from
inserting extra bids that increase the price charged from the winner. The possibility of
such manipulation decreases the trustworthiness and attractiveness of the second-price
auction. The Vickrey auction also requires adequate competition (as does the English
auction, though); otherwise the price paid by the winning bidder can be much too low
from the perspective of the seller. McMillan (1994) illustrates the problem of too little
competition with an example form a spectrum auction in New Zealand. In that
23
auction the winner with a bid of NZD 7 million only paid NZD 5000, which was the
second-highest bid.
The strongest argument for the English auction, however, is that made by Milgrom
and Weber (1982). They show that whenever there is a common value element in the
auction, the English auction provides the highest expected revenue. It is reasonable to
assume that in most auctions there is a common value element – the resale value of the
item, if nothing else – and thus it is only natural that the English auction is so
common.
Finally, an additional explanation for the popularity of the English auction, which is
often omitted in auction literature, is the fact that the general public – the potential
participants – are familiar with it. Throughout the centuries the English auction has
been all but synonymous with the term auction, and it has become something of an
industry standard. Bidders are more prone to participate in an auction which they are
familiar with – especially the risk averse ones. Also, learning the rules of an auction
always takes time and constitutes a transaction cost. Therefore, bidders tend to choose
auctions for which they already know the rules. Thus it is in the interest of the
auctioneer to organize such an auction, because the more there are participants the
higher the expected revenue.
2.5 Optimal Auctions
The term “optimal auction” can either refer to an efficient auction (as understood by
Vickrey 1961), or to a revenue maximizing auction (as understood by Myerson, 1981,
and Riley and Samuelson, 1981). An efficient auction maximizes the welfare of the
society as a whole, where as a revenue-maximizing auction maximizes the payoff to the
seller. The four basic auction mechanisms described above are efficient, but not
necessarily revenue maximizing. In this section I will discuss the design of a revenue
maximizing auction. Also, in this thesis, the term “optimal auction” refers to a revenue
maximizing auction.
Conceptually, the derivation of an optimal auction mechanism is straightforward: the
task is to simply maximize the seller’s expected revenue subject to individual rationality
and incentive compatibility constraints. In practice, however the problem becomes
24
easily intractable unless we resort to restrictive assumptions such as the IPV framework.
In the IPV framework it can be shown that all the four basic auction mechanisms are
optimal, if an appropriate reservation price is added (Myerson, 1981, Riley and
Samuelson, 1981). The optimal reservation price is set to mimic the expected bid from
the bidder with the second highest valuation, and it is strictly greater than the seller’s
own valuation for the item. The reasoning behind this is that if the reservation price is
higher than the valuation of the second highest bidder, the seller’s revenue is
increased, because the winner now has to pay the reservation price, and not the second
highest bid. However, it is also possible that the reservation price exceeds the valuation
of the highest bidder as well, and no sale takes place even though the valuation of the
seller was lower than that of the highest bidder. Thus, the optimal auction is not always
efficient. If designing a real auction, one should consider the results of these theoretical
models with caution, though. For instance, the number of bidders is treated as fixed
and exogenous in the models. In reality, bidder entry is endogenous, that is dependent
on the auction rules, and opening prices among others. Bajari and Hortaçu (2003) find
out that when a secret reservation price is determined by the seller, fewer bidders
entered the auction, which on average resulted in lower revenues for the seller.
Relaxing the IPV assumptions results in very complex optimal mechanisms. For
instance, if the bidders are asymmetric, then it is optimal to favor the low-valuation
bidders, because it forces the high-valuation bidders to bid higher than what they
normally would. On the other hand, in the case of risk averse bidders it is optimal to
subsidize high bidders who lose and penalize low bidders. The optimal auctions are
very complex, and almost impossible to implement in practice, because the design of
the optimal mechanism requires a lot of information about the bidders’ valuations, risk
attitudes etc. Efficient auction mechanisms are not as complex as optimal mechanisms,
and they are better for the society as a whole. Hence, in the remainder of this thesis,
efficiency will be used as a primary measure of the goodness of any auction
mechanism.
25
2.6 Empirical Studies of Auctions
Theoretical studies of single-item auctions abound, as is evident from the previous
sections. Another strand of economics is interested in the practical side of auctions. A
key question is how realistic the theoretical models are (or how realistic the
assumptions behind the models are). On the one hand researchers have studied real
auctions, and on the other hand they have conducted controlled experiments to study
human behavior in different auction settings.
2.6.1 Field Studies
Field studies of auctions observe bidding behavior in real auctions and try to test the
predictions of auction theory. The problem with empirical analysis is that any tests
should be able to first determine underlying risk preferences of the bidders, the
independence (or interdependence) of bidders’ valuations, and the symmetry of the
bidders. The studies usually indicate that the theory does not hold in practice, but the
results are contestable more often than not. Some field studies conducted in online
auctions study the behavior of the bidders (see e.g. Bapna, Goes and Gupta, 2000,
2003). Among the most studied and documented real-life auctions are the radio
spectrum license auctions held by the Federal Communications Commission (FCC)
in the United States. The FCC auctions will be discussed further in section 3.1.2.2.
2.6.2 Experimental Studies
It is a bit difficult to analyze empirical data on auctions in the light of auction theory,
because so many parameters (e.g. valuations, risk attitudes) are difficult to observe and
to control. The school of experimental economics founded by Vernon Smith, Charles
Plott and John Ledyard has attempted to test the theories in practice with human
subjects, but under more controlled circumstances than what is possible in field
studies. The experimental studies try to cut a balance between the realistic nature of
the bidding situation and full control of the parameters.
IPV model
Numerous experimental tests of the IPV model have been conducted. Kagel (1995)
reviews in detail experiments on all aspects of the traditional auction models. Bichler
26
(2001) gives a more general overview. The most commonly tested issues are the
revenue equivalence of the four basic auction mechanisms and the efficiency of the
mechanisms. Most research has been limited to the pairwise comparison of the Dutch
and first-price auctions, and the English and Vickrey auctions. The revenue
equivalence has not held even between the allegedly strategically equivalent auction
pairs, so there has been no need to test the equivalence of all four mechanisms.
The experiments reviewed by Kagel (1995) conclude that subjects do not behave in
strategically equivalent ways in first-price and Dutch auctions or in Vickrey and
English auctions. The revenue equivalence between the auctions did not hold either.
Equilibrium prices were consistently higher in first-price sealed-bid auctions than in
Dutch auctions (Coppinger, Smith and Titus, 1980, Cox, Roberson and Smith, 1982).
Similarly, prices in Vickrey auctions exceeded those of English auctions (Kagel,
Harstad and Levin, 1987). Interestingly, the bids in Vickrey auctions consistently
exceed the dominant strategy. This is probably due to the fact that the dominant
strategy in Vickrey auctions is far from obvious.
Lucking-Reiley (1999) presents a more recent test of the revenue equivalence, which
was conducted in the WWW-environment. He auctioned off collectable Magic cards
worth around $2,000 by posting advertisements to news groups and using e-mail as a
communication tool. His experiment differed from previous experiments in the sense
that the bidders did not know they were participating in an experiment. On the one
hand, this made the setting more authentic but on the other hand, it made it
impossible to control for the assumptions underlying the IPV model. It was also
impossible to control the number of bidders in each auction. Lucking-Reiley
concludes that the equilibrium prices in the Dutch auctions were significantly higher
than in the first-price auctions. The prices in the Vickrey and English auctions were
about the same, although bid-level data indicated some tendency for bidders to bid
higher in the English auctions. Interestingly, these results conflict with those reviewed
by Kagel (1995). Lucking-Reiley explains the higher prices in the Dutch auction with
the fact that the number of bidders in the Dutch auctions was on average higher than
in the first-price auctions. A possible explanation for the higher than expected revenues
in the English auctions is affiliation of bidders’ valuations.
27
The results of the experiments indicate clearly that the revenue equivalence between
auction mechanisms does not hold outside the ivory tower. Explanations for the
experimental results have been sought in the restrictive and unrealistic nature of the
IPV assumptions. The experimental tests of the efficiency of different auction
mechanisms give further reason to doubt the realistic nature of the assumptions behind
the IPV model.
Common Value Model
Most experiments in the context of the Common Value Model focus on the existence
of the winner’s curse. Kagel (1995) reviews numerous studies on the winner’s curse.
Experiments show that especially inexperienced bidders suffer from the curse
(Bazerman and Samuelson, 1983, Kagel and Levin, 1986). With enough experience,
though, bidders learn to bid below their estimates and are able to obtain profits from
auctions (Garvin and Kagel, 1994). The experiments conducted by Levin, Kagel and
Richard (1996) show that the English auction increases the expected revenue for the
bid taker, as predicted by theory. However, if bidders make a mistake and do not take
the winner’s curse into consideration in sealed-bid auctions, they might bid higher
than in the English auction (and end with a negative profit).
2.7 Criticism of the Traditional Single-Unit Models
Rothkopf and Harstad (1994) criticize the single-unit models presented in literature.
Most of these models apply a game theoretic approach to auction design. Rothkopf and
Harstad claim that this approach is too simplistic to accurately reflect real world
auction situations. For instance, the models assume that the auction occurs in isolation
of all previous and future auctions. This is not true in reality. Auction participants
cannot optimize their behavior with respect to one single auction; they have to think
about their reputation that affects the outcomes of future auctions. Also, the
assumptions made about the bidders are too restrictive and do not reflect reality. The
bidders are not symmetric, and they may have different risk attitudes. Bapna, Goes and
Gupta (2000) studied online auctions and identified three different bidder types, each
having a different risk attitude and bidding strategy. The results of numerous
experiments reviewed in section 2.6 have demonstrated that the traditional single unit
28
models do not reflect reality. The intricate optimal auction mechanisms presented in
section 2.5 seem to be too complex to be implemented in practice. Rothkopf and
Harstad’s critique lies heavily on their observation that no bidder seems to be using a
game-theoretic model to decide how much to bid. Rothkopf and Harstad’s critique
indicated a new direction for auction research, and since the 1990s the focus of
auction research has been shifting from purely theoretical treatments of auction
models toward more practice oriented studies.
2.8 Multiple-Unit Auctions
The first attempt to make auction models more realistic was the inclusion of multiple
units of the same item in the auction. These units could be auctioned either
sequentially (one at a time) or all at the same time. Using a sequential auction is also
the simplest way to model interdependencies among auctions, and the effect of
reputation. Because of the multiple units, the auctioneer has to decide whether the
pricing scheme is discriminatory or competitive (non-discriminatory). Discriminatory
pricing means that each winning bidder pays the price equal to her bid (i.e. all winners
end up paying a different price for the units). In competitive pricing, all the winners
pay the price equal to the lowest winning bid – or the price equal to the highest losing
bid.
In the simplest version of the multiple-unit auction, each bidder only has use for one
unit. This reduces the auction to a price-only situation, which is only slightly different
from the single-unit case. All the four basic mechanisms extend to this setting easily. In
the more complicated case, the bidders can bid for any number of units for sale. Here
the bids bi = (pi, qi) are vectors with two components, one indicating the per-unit price
(pi) and the other indicating the desired quantity (qi). The first-price, second-price and
English auctions have their extensions here too, but the Dutch auction is not suitable
for the general multiple-unit auction.
The general multiple-unit extension of the first-price auction is the pay-your-bid
auction, where bidders provide a “demand schedule”, that is, a price for each unit they
are interested in. Each bidder receives the number of items she demands for the
clearing price, and pays according to her bids. The general multiple-unit extension of
29
the Vickrey auction is quite complicated. Bidders submit demand schedules, and the
winners are the ones with the highest bids. However, the payment made by the bidder
on the jth unit she wins is equal to the jth highest rejected bid of her opponents. A
simple example adopted from Maasland and Onderstal (2006) will explain it clearly.
Assume there are three bidders and three units for sale. The bids are presented in the
table below:
Table 2 Example of a multiple-unit Vickrey auction: bid prices for the bidders
Bidder 1 Bidder 2 Bidder 3
1st unit 10 8 6 2nd unit 9 4 3 3rd unit 7 3 3
Bidder 1 wins two units and Bidder 2 one unit. The highest losing bid from Bidder 1’s
competitors is p = 6 from Bidder 3, so Bidder 1 pays 6 for the first unit. The second
highest losing bid from her competitors is p = 4 from Bidder 2, so the total payments
from Bidder 1 are 10. Bidder 2’s payment is 7, which is the highest losing bid made by
her competitors.
It is sometimes mistakenly thought that the uniform-price auction is the multiple-unit
extension of the Vickrey auction (Maasland and Onderstal, 2006). However, this is not
the case, and actually the uniform auction is not even efficient, as the Vickrey auctions
are reputed to be. The uniform-price auction is similar to the pay-your-bid auction,
except that each bidder pays the same price for each unit, and the price is equal to the
highest losing bid. Because it is assumed that bidders bid for more than one unit, it is
possible that the highest losing bid is made by one of the winning bidders. Thus, there
is the possibility that a bidder can affect the price she has to pay, and bidding your
valuation no longer is a dominant strategy.
The multiple-unit extension of the English auction is the Ausubel auction (Ausubel,
2004). In the Ausubel auction the price starts from zero, and increases continuously. At
each price level, the bidders announce the quantity they would be willing to purchase.
Then, each bidder’s demand in turn is taken out of the total demand. If the demand
from all the other bidders exceeds supply, nothing happens, and the price is increased
to the next level. However, if there exists a bidder, without whom the total demand
30
would be smaller than the supply, this bidder is awarded the “shortage” for the current
price. The awarded units are removed from the supply, and price is increased to the
next level, and the auction continues until all the units have been sold.
Multiple-unit auctions have received a lot of attention (see the above-mentioned
references and e.g. Vickrey, 1961, Wilson, 1979, Engelbrecht-Wiggans, 1988, Maskin
and Riley, 1989 and Tenorio, 1999), but the research in the area has followed the
tradition of the single-unit models. The auction models are usually based on traditional
IPV assumptions and the goal is to obtain tractability and equilibrium strategies.
However, despite the restrictive nature of the studies, the multiple-unit auction models
with bid vectors, albeit consisting of only two components, paved the road for the study
of multiple-item (see Chapter 3) and multi-attribute models (see Chapter 4).
31
3 MULTIPLE-ITEM AUCTIONS – AN INTERDISCIPLINARY
FIELD
Where the traditional auction research is dominated by economists, the study of
multiple-item auctions has been of interest to computer scientists, operations
researchers and decision analysts as well as economists and game theorists. Most
attention has been given to combinatorial auctions, which are multiple-item auctions
in which package bidding is allowed. Multiple-item auctions, especially combinatorial
auctions, are so complex to organize that it is almost impossible to do it without the aid
of computers. Thus the research in combinatorial auctions has advanced along with
the development of computers. The article by Rassenti, Smith and Bulfin (1982) is
reputed to be among the first in the field of combinatorial auctions. Since then
computers have developed greatly, and the invention of the Internet has lowered the
threshold to organize all kinds of electronic auctions. Since combinatorial auctions are
very complex they also provide a fruitful ground for a lot of different kinds of research.
Thus, in the past decade or so combinatorial auctions have become a hot topic in
auction research. Researchers with different backgrounds have all found a perspective
on combinatorial auctions they can contribute to. Economists and game theorists
construct mechanisms with theoretically desirable properties, such as efficiency.
Operations researchers study the integer programming aspects of winner determination
and feedback mechanisms, and computer scientists create algorithms for the winner
determination problem. Decision analysts have been interested in the creation of bids,
and developing tools to help bidders evaluate their preferences over bundles. In what
follows I will present research from different fields, and show how they complement
each other.
In this thesis I will not consider single-unit and multi-unit combinatorial auctions
separately, because their properties are essentially the same. Our research assumes
multiple units, but the single-unit case can be dealt with as a special case. Most
research considers the single-unit case, and hence part of the following discussion is
from the viewpoint of single-unit multiple-item auctions.
32
At first sight, the step from single-item, multiple-unit auctions to multiple-item
auctions does not seem that big. However, the difference in complexity can be
enormous. The complexity arises from the underlying assumption that the reason for
holding a multiple-item auction is that bidders have nonlinear preferences over
bundles of items. In other words, the items are either substitutes or complements, so
that the value from a bundle is not the sum of the values of its components. For
complements it holds that v(A, B) ≥ v(A) + v(B), and for substitutes v(A, B) ≤
v(A) + v(B). Examples of such preferences abound. In radio license auctions there may
be synergies in obtaining licenses for adjacent areas, or two licenses of different
frequency for the same area can be substitutes (Pekeč and Rothkopf, 2003). Also many
reverse auctions exhibit nonlinear preferences. E.g. it is easy to imagine that
transportation services (trucking) have complementarities: the cost per haul decreases,
if the truck is full on all routes. In the reverse setting, complementarities between items
translate to a subadditive cost function c(A, B) ≤ c(A) + c(B). Nonlinear preferences,
such as the ones described above, make bidding a complex task, as the value of one
item to the bidder is dependent on what other items she wins. Also, the task of
determining the winners of the auction – which so far has been relatively
straightforward – can become difficult.
The auction mechanism the auctioneer chooses to use has a major effect on the
complexity of bidding and winner determination. On a macro level, the choice is
essentially between allowing bids on combinations or not. Not allowing bids on
combinations makes the auction a lot easier for the auctioneer to handle, and there are
no problems with winner determination. The downside is that bidding is very difficult
and the outcome can easily be inefficient. Allowing combinatorial bidding enables the
bidders to better express their preferences, but winner determination becomes
cumbersome. Bidding is still difficult, but for different reasons. In the following I will
briefly discuss some non-combinatorial multiple-item mechanisms before going into a
detailed discussion on combinatorial auction research.
33
3.1 Non-Combinatorial Auction Mechanisms
If bids can be made on single items only, the multiple-item auction is essentially a
collection of several single-item auctions. These auctions can then be held sequentially
or in parallel.
3.1.1 Sequential Auctions
In a sequential auction, each item in the bundle is auctioned one at a time, one after
another. This design is simple for the auction owner, since it is easy to define the
winner in each auction. The winner is simply the bidder with the highest bid. For the
bidders, however, the sequential design imposes grave difficulties. A bidder’s valuation
of each item depends on what other items she wins in the ensuing auctions. Therefore,
in order to establish her optimal strategy in one auction, she will have to try to guess
the outcomes of the future auctions. This involves speculation of the possible strategies
of the competitors, which in turn depend on the outcome of the auction at hand. The
computational costs are high, and in auctions with relatively large numbers of items
and bidders, the calculation of the optimal strategy becomes intractable. Hence, the
outcomes of the auctions are easily inefficient: the bidders do not obtain combinations
they wanted, or pay more than they would have wanted for the combinations they do
get. This problem is referred to as the exposure problem (Rothkopf, Pekeč and Harstad,
1998, Pekeč and Rothkopf, 2003).
3.1.2 Simultaneous Auctions
An alternative to a sequential auction is a simultaneous auction. In this auction, the
items are auctioned in separate auctions that run at the same time. Here I will first
discuss the general properties of simultaneous auctions, and some improvement
suggestions to the design. After that I will present the Federal Communications
Commission’s (FCC) radio spectrum license auctions as an example of simultaneous
auctions. The FCC auctions have received a lot of attention in literature due to their
large size, and the fact that there are clearly complementarities and substitutabilities
between the items.
34
3.1.2.1 General Properties of Simultaneous Auctions
The exposure problem still plagues the bidders in simultaneous auctions, but not as
badly as in sequential auctions. The determination of the winner in each auction is still
as simple as in sequential auctions, but the bidders’ task has become a little easier. The
bidders can observe each other’s bidding behavior in all the auctions5, which reduces
the need for speculation. Ledyard, Porter and Rangel (1997) compared simultaneous
and sequential auctions, and concluded that simultaneous auctions are more efficient.
However, some problems remain. The bidders still do not know which items they will
receive when all the auctions are closed. Hence, they cannot determine their
valuations for the items a priori and it is impossible to establish the optimal bidding
strategy. Moreover, in simultaneous auctions each bidder would like to wait until the
end to see what the going prices for the items will be, and then optimize her own bids
taking the final prices into consideration. Because all bidders would prefer waiting, no
bidding would begin in the first place. So called activity rules could be established to
guarantee bidding (McAfee and McMillan, 1996). This means that each bidder must
bid at least a certain volume by predetermined points in time, or her future bidding
rights are reduced. The activity rules are sometimes referred to as Milgrom – Wilson
activity rules (see Milgrom, 1998) after their developers.
Sandholm (2000) proposes some methods to improve the efficiency of sequential and
simultaneous auctions. One approach would be to establish an after market where the
bidders can exchange items once the auction has closed. This reduces the inefficiency
of the auction outcome, but may require an impractically large number of exchanges.
Another, more practical approach would be to allow bidders retract their bids. In this
case it is important to guarantee that retractions do not diminish the auctioneer’s
payoff. There are many ways to take care of that. For example, if the closing price is less
than the retracted bid, the bidder who retracted her bid has to pay the difference
(McAfee and McMillan, 1996). Sandholm (2000) suggests a leveled commitment
protocol be used. In this protocol, the penalty from retracting a bid is set up front, and
5 This is true only if the auctions are held in the open-cry format. Simultaneous sealed-bid auctions are equivalent to sequential auctions, because no additional information can be obtained.
35
also the auctioneer is allowed to decommit from the auction outcome. This reduces
the risk for the bidders, as the penalty from retracting is known in advance.
3.1.2.2 The FCC Spectrum License Auctions
The multiple-item auction that has received the most attention in the past decade is
the Federal Communications Commission’s (FCC) spectrum license auctions held
since the mid-1990s. McMillan (1994) provides a thorough description of the auction
process. The items for sale were regional radio spectrum licenses covering the
wavelengths used for personal communications services (PSC), such as cellular phones
and wireless computer networks. What made the auctions so unique were their size
and complexity. Thousands of licenses were for sale, bidders were many and diverse,
ranging from large national telecommunications companies to small local firms, and
the estimated value of the 1994 auction was over $10 billion. Dozens of economists
were hired by the telecommunications companies and the FCC to help design the
auctions.
The complexity of the auctions was increased by the realization that there were
potentially complementarities between the licenses. The potential efficiencies derived
from the aggregation of licenses have both engineering and economics aspects. First,
the fixed costs of technology acquisition and building up a customer base can be
spread over several licenses. Second, there are often problems of interference at the
boundaries of license areas so it is cost-efficient to operate in adjacent areas. Third,
consumers will value the ability to use the same phone when traveling all over the
country. The main question in the auction design became then how to best take the
complementarities into consideration without compromising the functionality of the
mechanism. Also, because the seller was the government, revenue maximization was
not the primary goal, but rather the efficient allocation of licenses.
The FCC decided to run simultaneous multiple-round auctions (SMA or SMR)
developed by Paul Milgrom, Paul Wilson and Preston McAfee (as documented by
Milgrom, 1998 and 2000). Because the bidders were informed of the competing bids
after each round, this format was close to an open-cry auction. The bidders were also
allowed to withdraw bids, but if the equilibrium price ended up being lower than the
36
retracted bid, the bidder was obliged to pay the difference. The parallel auctioning of
the licenses combined with the option to retract bids gave flexibility for bidders to
aggregate licenses. Moreover, this way the bidders were able to switch to their “back-
up” combination, if their most preferred one turned out to be too expensive. The FCC
contemplated allowing combinatorial bids, i.e. bids for a combination of licenses. This
would have allowed the bidders to express their synergies over aggregation of licenses
directly in monetary terms. Theoretically, this could have produced more efficient
results than a parallel auction. However, the FCC was afraid that administrative or
computer breakdowns would occur due to the computational complexity imbedded in
combinatorial bidding. Allowing combinatorial bids could make the auction too
complicated for the bidders causing the complexity costs to outweigh the potential
efficiency gains. Also, the threshold problem creates incentives to free ride (see section
3.2.1.3 for a full explanation), which was seen as relevant problem impeding efficient
outcomes. Thus, the FCC decided against combinatorial bidding.
Very recently, the FCC experimented with allowing package bidding in one of the
spectrum license auctions (Auction #73 of the 700 MHz band). Combinatorial bidding
was allowed in one license block containing 12 licenses (FCC, 2007). However, the
allowed combinations were restricted to three packages: “50 states” (licenses 1-8),
“Atlantic” (licenses 10 and 12) and “Pacific” (licenses 9 and 11). At the end of the
auction, only the bid on the “Pacific” package was among the winners; all other
licenses were sold individually (FCC, 2008).
3.2 Combinatorial Auctions
Combinatorial auctions are defined as auctions in which multiple but different items
are sold, and bidders are allowed to make indivisible bids on packages (Pekeč and
Rothkopf, 2003). Bids are vectors (q1, …, qK, p), where the first K elements indicate the
quantities for the items, and the last element indicates the price for the whole package.
In single-unit combinatorial auctions the qi’s simply indicate whether a particular item
is in the package or not. Indivisibility refers to the restriction that all bids have to be
accepted as a whole or not at all; no partial bids can be accepted. Already Rassenti et al.
(1982) acknowledged that allowing combinatorial bids alleviated many of the problems
37
in sequential and simultaneous auctions. For instance, there is no need to estimate the
opponents’ strategies in other auctions (possibly later in time) when all items are sold
in one auction. Researchers also argue that combinatorial bidding allows the bidders to
better express their preferences over bundles, and therefore the auction outcome
should be more efficient than in non-combinatorial auctions. Ledyard, Porter and
Rangel (1997) ran a series of laboratory experiments to compare combinatorial
auctions to sequential and simultaneous auctions. They concluded that combinatorial
auctions are more efficient (they produce outcomes closer to the efficient allocation)
than sequential or simultaneous auctions. Banks et al. (2003) compared the
simultaneous multiple-round auction (SMA) used by the FCC to a combinatorial
auction, and reached similar results.
3.2.1 Challenges with Combinatorial Auctions
Despite all the theoretical benefits accruing from combinatorial bidding,
combinatorial auctions have not been used that much in practice. The reasons for this
arise from three properties that distinguish combinatorial auctions from other auction
types: complexity of winner determination, complexity of bid formulation, and the
strategic gaming element, which leads to what is known in the literature as the
threshold problem (Pekeč and Rothkopf, 2003).
3.2.1.1 Complexity of Winner Determination
The winners of a combinatorial auction are the bids that maximize the bid taker’s
revenue (or minimize the cost) and allocate each item to only one bidder. If there are
multiple units of each item, the number of units allocated cannot exceed the number
of units available. All bids are assumed indivisible, also called all-or-nothing bids. Thus,
the solution to the auction is found from a set of disjoint bids, which maximizes the
seller’s revenue.
The winner determination problem (WDP) can be formulated as an integer
programming (IP) problem. Because most applications of combinatorial auctions are
in procurement situations, I will now present the WDP in the reverse auction setting.
38
Consider K items and assume that dk (k = 1, … , K) units (nonnegative integers) of each
of K items are requested by the buyer, defining demand. Now each bid j by bidder i is a
(K+1)-dimensional vector: (qij1, qij2, … , qijK, pij), where 0 ≤ qijk ≤ dk are nonnegative
integers (quantities of item k) and pij (price of the bundle) is also a real positive
number. In other words, bidder i’s jth bid is an offer to deliver qijk units of each item k
for a total price of pij.
The WDP determining the status of each bid by each bidder at any given moment in
the auction is formulated as an IP problem as follows:
{ }1,0
,...,1..
min
1 1
1 1
∈
=∀≥∑∑
∑∑
= =
= =
ij
k
N
i
n
j
ijkij
N
i
n
j
ijij
x
Kkdqxts
px
i
i
(2)
The variable xij indicates whether bidder i’s jth bid is among the winners (xij = 1) or not
(xij = 0), ni is the number of bids placed by bidder i, where i = 1, …, N . In case there is
only one unit of each item, dk = 1 for all k, and { }1,0∈ijkq depending on whether item
k is in the bid xij or not. This formulation allows any number of bids per bidder to be
among the winners. In some auctions the auction owner may want to limit the number
of winning bids per bidder to one. In that case, a constraint ixin
j
ij ∀≤∑=
,11
should be
added.
The problem of finding the optimal outcome in a reverse combinatorial auction is
equivalent to a set-covering problem (SCP). The SCP is a close relative of the set-
packing problem (SPP) of the forward auction (de Vries and Vohra, 2003). The set-
packing problem is known to be NP-complete (Rothkopf, Pekeč and Harstad, 1998),
which means that as such there is no revenue maximizing algorithm that can solve the
problem in polynomial time6. The properties of the SCP and SPP are a little different,
e.g. the SCP can be a little easier to approximate, but the most important results
6 In the analysis of computational complexity, time is measured in the number of computations required for solving a problem.
39
regarding the computational manageability of the problems are very similar (de Vries
and Vohra, 2003).
It is customary to study the “worst-case scenarios” for the time needed by any algorithm
to compute a solution. Mathematicians have focused on how the computation time
grows as a function of the input. A class of computational problems is labeled
“computationally manageable”, if an upper bound on computation time for all
problems can be expressed as a polynomial function of the inputs. The notation f(n) =
O(g(n)) means that for a function f(n), which is the number of calculations required to
solve a problem whose input size is n, there exists a limiting function g(n) so that
f(n) ≤ cg(n), where c is some constant, when n grows large (see Papadimitriou and
Steiglitz, 1982, p.159). For example, in non-combinatorial auctions, where the winner
determination can be solved by picking the highest bidders for each item separately,
the auction can be solved in O(NK) time, N represents the number of bidders and K
the number of items. So clearly, the non-combinatorial auction can be solved in
polynomial time no matter how large the input. The number of possible combinations
in a combinatorial auction is 2K-1 and the winner determination is solved in O(KK)
time, so there is no polynomial function that would express the number of
computations required to solve the problem. This also means that the exhaustive
enumeration of all possible outcomes is not a viable method for searching the optimal
allocation, unless the number of items is very small. The general WDP can thus be
declared computationally unmanageable.
3.2.1.2 Complexity of Bid Formulation and Communication of Preferences
Combinatorial bidding allows bidders to express their preferences over the different
items in the auction. However, this can be a complex task, because there are 2K-1
combinations (or more, if there are multiple units) over which the bidder should be
able to express her preferences. As K grows, it is impractical, not to mention time
consuming, to evaluate all conceivable combinations. Somehow bidders should be
able to identify, which combinations are interesting to them, and concentrate on
evaluating those combinations. Hoffman, Menon and van den Heever (2004) and
Jones and Koehler (2002) argue that bidders do not even think in terms of the items
40
they want in the bundle, but have other objectives which can be achieved through
different combinations. For instance, in the FCC radio license auctions bidders
desired a certain level of population coverage and bandwidth (Hoffman et al., 2004),
and not necessarily a particular license. Similarly, in auctions for airtime for TV
advertisements the bidders are essentially interested in acquiring a large exposure
among a particular demographic group, and obtaining particular slots is simply a
means to achieving the objective (Jones and Koehler, 2002). Thus, the bidders need to
translate their objectives and constraints (e.g. budget) into bid combinations. They also
need to estimate the value of each bundle to them in order to attach prices to the bids,
which is not necessarily a trivial task. In fact, An, Elmaghraby and Keskinocak (2005)
report that in a combinatorial auction for transportation services most bidders abstained
from placing combinatorial bids. A plausible explanation for this phenomenon is that
bidders found the construction of combinatorial bids too difficult.
When the bidders have managed to construct a set of combinations they would like to
bid on, they need to communicate their preferences to the bid taker. Usually auction
owners have defined a specific bidding language that has to be used to encode bids and
preferences. A bidding language both defines the exact syntax to be used in submitting
bids, and defines what kind of interdependencies can be expressed between bundles.
For instance, some bidding languages allow logical operators like “and”, “or” and
“not”, but some do not. Thus, it depends on the language, how well the bidders can
express their preferences. If the language is not fully expressive, an exposure problem
similar to the one in sequential and simultaneous auctions can still occur (Pekeč and
Rothkopf, 2003). Bidders may want to place bids on many different combinations in
hope of winning at least something, but they might not want to win all of them (e.g.
they may not have enough capacity to produce everything). A bidder would then like to
express her bid in the form “either combination A or B, but not both”. More primitive
bidding languages (OR, XOR) do not allow bidders to make complicated bids that
would allow bidders to express different preferences over a set of bundles. OR bidding
language does not allow bidders to restrict the number of bids that might become
winners. Basically, in OR language, any number of the disjoint bids placed by the
bidder could become winners. XOR language is the other extreme. There at most one
41
of the bidder’s bids can be among the winners, so it requires the bidder to
communicate every single combination she might be interested in. More advanced
languages, such as combinations of OR and XOR languages (OR-of-XORs and XOR-of-
ORs), enable more expressive bidding, and OR* is a compact and expressive language.
Nisan (2006) provides an extensive review on different bidding languages and their
properties.
The catch is that usually, the more expressive the bidding language is, the more
difficult it is to use, and the more complicated the computations usually get. Thus the
solution of the WDP slows down even further (Nisan 2000). However, there are
exceptions to this rule; e.g. XOR is easier to compute than OR. One way to circumvent
the use of complicated bidding languages is to use dummy items in bids. A dummy
item is an item that costs nothing, so adding it to two otherwise disjoint bids makes
them overlapping without any added costs. The artificial overlap created by the
dummy item ensures that the two bids cannot both be among the winners (Fujishima
et al. 1999). Even if there were no dummy items available, researchers have observed
that bidders used the cheapest items in the auction as dummies.
3.2.1.3 Strategic Gaming in Combinatorial Auctions: the Threshold Problem
Even if the bidder has managed to sort out her preferences, placing bids in the actual
auction is not simple. This is because in combinatorial auctions there can be multiple
winners. In order to become a winner with other bidders, the bidder’s bid has to
complement the other bidders’ bids. A phenomenon called the threshold problem is
identified in literature (see e.g. Pekeč and Rothkopf, 2003). The threshold problem
refers to the situation when small, “local” bidders bidding on single items cannot beat
alone a currently winning bid on the whole bundle made by a “global” bidder. This
leads to gaming between the bidders, as they all try to maximize their profit but still be
among the winners. Consider the following simple example of a reverse auction. There
are four bidders (a, b, c and d) bidding on three items (x1, x2 and x3), and the demand is
one unit for each item. The bids submitted by the bidders are ba(x1) = bb(x2) = bc(x3) = 5
and bd(x1, x2, x3) = 13.5. Assume that the costs for each bidder for each individual item
ci(xj) = 4. Bidder d is currently the winner, because her bid price 13.5 is lower than the
42
combined price of 15 offered by the other bidders. Bidders a, b and c could each afford
to lower their bid price by one unit making the total cost 12, which would allow them
to become provisional winners. In fact, it would be enough that two of them lowered
their bid. However, none of them alone could afford to lower the price so much that
the combined total cost would go below the bid of bidder d. The bidders would have to
somehow come to a mutual agreement to lower their bids in order to oust the current
winner. In this example, there is also the problem of potential free riding by one of the
bidders. Since it is enough that only two bidders lower their bids from 5 to 4, every
bidder hopes to be the one who can become a winner without having to reduce the
price. Thus, it is very likely that none of them reduce their price, and the auction
outcome is not efficient.
The threshold problem extends to any situation in which a number of bidders together
are trying to beat a bid on a larger combination. Chances are that more than one
bidder needs to adjust their bid price, and that the bidders have the incentive to free
ride. A partial solution to the threshold problem would be to accept bids that do not
become winners at that point in time in the auction (Pekeč and Rothkopf, 2003). This
could make the design of activity rules more difficult, because it will become harder to
distinguish between serious bids and attempts to merely fulfill activity rules without
having to compromise on profits (see section 5.2 on auction rules). The threshold
problem becomes even more difficult to overcome, if the auction is a sealed-bid
auction. In that case, the bidders do not even know the prices in the other bids, which
they are trying to coordinate with.
3.2.2 How to alleviate the problems?
A big part of combinatorial auction literature concerns alleviating one or several of the
above-mentioned problems. Computer scientists and operations researchers have
tackled the computational issues, and decision analysts and operations researchers have
studied ways to support bidders in valuing bundles and placing bids.
43
3.2.2.1 Alleviating Computational Problems
There have been two different approaches offered as solutions to computational
problems. The first approach aims at developing fast algorithms, which would enable
the solution of larger auctions quickly. Also some approximation algorithms have been
suggested. The second approach to the WDP is to restrict the bid space (the
combinations that can be bid on or the number of combinatorial bids) so that
computational manageability can be assured.
Exact and Approximation Algorithms for WDP
An exact algorithm is an algorithm which guarantees an optimal solution. Several exact
algorithms, which use a variety of techniques, have been developed. Some algorithms
utilize integer programming, others prune the search tree, and some are based on
dynamic programming.
Rothkopf, Pekeč and Harstad (1998) present a dynamic programming algorithm, which
makes it possible to solve the winner determination problem in O(3K) time, where K is
the number of items. The algorithm uses the observation that for each possible
combination S of the items, the maximal revenue comes either from a single bid b(S)
or from the maximal revenues of two disjoint exhaustive subsets of S. The algorithm
starts from singletons and proceeds systematically to larger sets until it reaches M, the
combination containing all items. The advantage of the algorithm is that it calculates
the revenue maximizing solutions for the subsets only once each time the winner
determination problem is solved. The weakness of the dynamic algorithm is that it
makes the same number of calculations independent of the number of actual bids.
This is because it goes through every possible combination S even if there is no bid for
it. The algorithm enables the auction owner to determine the calculation time already
prior to the auction, but it cannot take advantage of the potentially small number of
actual bids. Due to this fact, the dynamic algorithm functions in the worst case only if
the number of items for sale is 20-30.
Branch-and-bound algorithm is a general algorithm that finds an optimal solution to
any integer programming problem. However, since essentially the branch-and-bound
algorithm is based on enumerating all feasible solutions (although organized as a
44
search tree), it can become slow when the size of the auction grows. Fujishima et al.
(1999) were among the first to consider an algorithm based on an intelligent pruning of
the search tree. Since then, several studies have been written on more efficient search
algorithms.
Gonen and Lehmann (2000) have developed a branch-and-bound type of algorithm to
solve the integer programming problem (the WDP). Their algorithm is a depth-first
search, which calculates the values for each branch in turn, all the while updating the
current best solution. To speed up the search, Gonen and Lehmann suggest that the
algorithm estimates an upper bound7 for the objective function that can be achieved
from each branch. The upper bound is then compared with the current best solution.
If the value of the objective function in the best solution so far is higher than the value
indicated by the upper bound, that particular branch can be “pruned”, i.e. excluded
from further consideration.
Also Sandholm (2000) utilizes the intelligent pruning of the search tree in his
algorithm. The key observation of Sandholm (2000) is that in larger auctions the bid
space is necessarily very sparsely populated. For example, if the number of items for
sale is 100, it would take longer than the life of the universe to bid on all the 2100-1
combinations, even if every person on earth placed a bid every second. Even in smaller
auctions there hardly ever is a bid for every conceivable combination. Sandholm
(2000) proposes an algorithm that takes advantage of this sparseness in the bid space.
The algorithm generates a tree where each path consists of a sequence of bids
organized based on the items in the bids. The path ends when all items have been
used. Each path represents a feasible allocation, the revenue of which can then be
compared with other allocations. The algorithm is implemented as depth-first search.
This enables the auction owner to find feasible allocations quickly. Also, the algorithm
keeps track of the best solution so far, so in case the algorithm has to be terminated
before all the paths have been generated, the best solution so far can be obtained. The
most significant difference between Sandholm’s algorithm and the dynamic
7 Gonen and Lehmann consider forward auctions, hence it is useful to define an upper bound for the value of the objective function in some subset of bids. Correspondingly, in reverse auctions it would be useful to obtain a lower bound.
45
programming algorithm is that Sandholm’s algorithm generates only the paths for
which there are actual bids. In the worst case, Sandholm’s algorithm takes O(mK) time,
where K is the number of items and m the number of bids ( ∑=
=N
i
inm1
using notation
from (2)), to find the optimal allocation.
Sandholm and Suri (2003) and Sandholm et al. (2005) improve the algorithm
proposed in Sandholm (2000 and 2002). Their major revelation is that it is more
efficient to branch on bids rather than on items (as was done in the earlier algorithm).
They develop a new branching method, BOB, and an algorithm, CABOB, to be used
in a combinatorial auction.
Also other improvements to a basic branch-and-bound algorithm have been suggested.
Ono, Nishiyama and Horiuchi (2003) have developed a method for iterative
combinatorial auctions that utilizes the previous solution of the WDP to increase the
speed of the search algorithm. Their method can be combined either with different
search algorithms. Mito and Fujita (2004) suggest a way to order bids so that once a
branch-and-bound algorithm is applied, it finds an optimal solution faster. Günlük et
al. (2005) on develop a solution algorithm based on “branch-and-price.” Instead of
operating with all variables (bids), branch-and-prices starts with a small subset of
variables, and through the dual of the WDP it formulates and solves a pricing problem,
which helps identify good variables to include in the solution. Günlük et al. (2005) test
four different branching rules, and conclude that branching on items is better than
branching on bids. Yang et al. (2009) suggest that regardless of the branching rule
used, the search process would become faster, if “nagging” were used. Nagging refers to
the parallelization of the search space, where portions of the search tree are distributed
to individual processors operating simultaneously. Thus, instead of working though the
search tree one branch at a time, the algorithm would work on several branches
simultaneously.
Andersson, Tenhunen and Ygge (2000) note that the winner determination problem
can also be formulated as a mixed integer programming problem. According to them,
this formulation can utilize standard algorithms and the problem can thus be solved
using commercial software. Andersson et al. (2000) test a software package called
46
CPLEX and they conclude that in most instances it performs very well and the
computation times are smaller than achieved with Sandholm’s (2000) intelligent
algorithm, and comparable to those obtained with Fujishima et al.’s algorithm.
According to Sandholm et al. (2005) their most recent algorithm, CABOB, is often
drastically faster than CPLEX, and rarely drastically slower.
There have also been efforts to find heuristics (Mito and Fujita, 2004, Jones and
Koehler, 2005, Guo et al., 2006, and Özer and Özturan, 2009) and approximation
algorithms (see Crescenzi and Kann, 2006 for a review) that would be polynomial and
produce a “reasonably good” result instead of an optimal one. However, for some
instances even the approximation can be difficult, and a good solution may not be
found (de Vries and Vohra, 2003). Another problem related to the use of
approximation algorithms in auctions is that it can compromise the perceived fairness
of the auction mechanism (Pekeč and Rothkopf, 2003). Even though the approximated
optimum is close to the true optimum revenue-wise, it can be comprised of a totally
different set of winning bids than the true optimum. Thus the bidders cannot be sure
of the fairness of such auctions, and auctioneers are hesitant to implement them in
practice.
Restricting Combinations
A different philosophy on the WDP is to not try to “bang one’s head against the wall”,
but to constrain the bid space so that computational manageability is guaranteed. What
this means is that the combinations bidders are allowed to bid on are decided prior to
the auction. This way the auction owner can limit the number of combinations to a
level which assures computational manageability. According to Rothkopf, Pekeč and
Harstad (1998), there are three instances when structure of the bid space is such that it
guarantees computational manageability: nested structures, cardinality-based structures
and geometry-based structures. These structures are useful, because it is possible to
identify situations in real life where these structures could be natural, and not limit the
bidders from expressing their valuations.
Nested structures take advantage of the situation where there are disjoint groups of
items with synergies within groups but none between groups. If this were the case, all
47
bids could be restricted to contain items from only one group, the exception being a
bid on all items in the auction. An example of this kind of a situation could be an
auction in which the assets are on the East and West Coasts of the USA. Assume that
there are no synergies to be obtained from mixing assets from both coasts. The
auctioneer could then limit the permitted combinations to include assets from only
one coast, the exception being a bid on all the grand combination, i.e. all assets in the
auction. The optimal outcome of the whole auction is then the union of the optimal
allocations of the subauctions of East and West Coast assets. According to Rothkopf et
al., the optimal outcome of any such nested auction can be determined in O(K2) time,
where K is the number of items in the auction.
On some occasions the auction owner might have an idea on how large combinations
the bidders would like to submit bids on. In these situations it would be advantageous
to use cardinality-based structures. For example, an apartment building is converted
into condominiums which are then auctioned off. It is unlikely that there would be any
synergies in buying two or three condominiums, but purchasing large enough a block
to gain voting control might be in the interests of some real estate company. The
auction owner could then restrict the allowable combinations to include singleton bids
and large combinations. If a large combination is defined as |S| > K/2, i.e. it has to
include more than half of all the items, then there can be only one such bid in the
optimal outcome. It is easy to see that this kind of an auction is computationally
manageable regardless of the number of items for sale. Cardinality-based structures can
also be used when synergies can be obtained from small sets. For example, if bids are
allowed only for pairs, the winner of the auction can always be found in polynomial
time.
Geometry-based structures are applicable in situations where there are synergies
between “neighboring” assets. The simplest of these structures is a line structure in
which all assets can be ordered and placed on a line. Rothkopf et al. (1998) give an
auction selling radio frequencies in the cities on the East Coast as a hypothetical
example of such a situation. The cities could be numbered from north to south and
ordered on a line. Allowing bids on only consecutive licenses ensures the
computational manageability of the auction. The two-dimensional extension of the
48
line structure to a plane of squares is computationally manageable only, if bidding is
allowed on combinations including only rows or columns. An example of a situation in
which the auctioneer could use a two-dimensional structure is a coin auction. Coin
collectors are usually interested in either coins of different values (pennies, nickels,
dimes) from the same year or the same coin (say nickel) from different years. If the
coins to be auctioned are organized in a matrix according to their value and year, then
permitting bids on single coins, rows and columns not only ensures the computational
manageability, but also enables the bidders to express their valuations. The two-
dimensional structure can be extended to k dimensions using the same logic.
It is appealing for the auction owner to predetermine allowable combinations to ensure
the computational manageability of the WDP. After all, to be successful, the auction
must be trustworthy. If the potential bidders cannot be sure that a winner can be found
or that the winners are those who really made the best offers, they might choose not to
enter the auction at all. However, there is a trade-off in restricting the bids. The
efficiency of the auction is compromised, i.e. the auction owner’s revenue and the
bidders’ utilities are not as large as they could be, if the bidders were allowed to bid
according to their true valuations. This is due to the risk that the auction owner may
lack the knowledge to be able to recognize all combinations that would be important
for the bidders. Therefore, Park and Rothkopf (2005) suggest the bidders be able to
determine the combinations that are to be bid on. The bidders are asked to submit a
list of bids for all single items and also a prioritized list of combinatorial bids. The
winner determination problem is solved iteratively. First only the bids on single items
are considered. As this auction is always computationally manageable, it guarantees a
lower limit for the solution. After the initial solution is obtained, the first combinatorial
bids on each bidder’s list will be considered in addition to the single bids, and the
winner determination problem is solved. If a solution is obtained in reasonable time,
the algorithm proceeds to the second bid on the lists, and so on. The auction ends
when all bids on all lists have been considered, or if at some point the solution to the
winner determination problem is not obtained in reasonable time.
Park and Rothkopf (2005) argue that their approach has two advantages. First, the
auction will be regarded as fair by the bidders since they can choose the combinations
49
they want to bid on. Second, most of the efficiency is likely to be captured as the
bidders can freely express any synergies they might have and the bidders can be
expected to bid for the most important combinations first. The problem with Park and
Rothkopf’s method is that it is basically a sealed-bid auction. Once the lists are
submitted, the bidders cannot go back and improve their offers. The bidders face the
same decision problem as in single item price-only situations: they have to weigh the
extra profit against the probability of winning without knowing the decisions of other
bidders. If the auction is a multiple-unit auction there may easily be inefficiencies due
to a mismatch in the item quantities. Also, the bidder can never know when
comprising the list of bids, how many of her bids will be considered before the auction
ends. All this, I think, makes bidding in the auction perhaps overwhelming for
inexperienced bidders who have not studied auction theory. It is thus possible, that the
outcome of the auction is not as efficient as Park and Rothkopf assume it to be.
3.2.2.2 Preference Elicitation and Bidder Support
The problems combinatorial bidding poses for the bidders are receiving increased
attention from researchers. Most research is focused on preference elicitation.
Researchers have produced tools to help bidders translate bidder’ goals and constraints
into bids, and attach appropriate prices to the interesting bundles. These ideas are
similar to those presented in the fields of decision making and decision aiding, and
they have close links to the multi-attribute utility theory (MAUT). The bids in a
combinatorial auction can be seen as expressing a multi-attribute utility function in
which each item is an attribute (Sandholm and Boutlier, 2006). For some reason, the
gaming elements in the auction are disregarded. For instance, the threshold problem is
rarely addressed directly.
Because most auction mechanisms considered in the combinatorial setting are
iterative, also most preference elicitation schemes are designed for iterative auctions.
An iterative auction alone with feedback on prices and provisional allocations can be
understood as a kind of bidder support. The iterative format can help the bidders
because bidders are no longer required to supply bids on all 2K-1 combinations.
Instead, they can only supply a few bids every round, and adjust their strategies
50
according to information feedback coming from the auction (Pekeč and Rothkopf,
2003). However, the iterative auctions as such do not offer any explicit support for
preference elicitation, and they are often framed as auction mechanisms rather than
support tools, hence I will discuss them later in section 3.2.3.2 on different iterative
mechanisms.
Conen and Sandholm (2001) propose a selective preference elicitation approach. They
argue that in an iterative auction the auctioneer does not have to elicit each bidder’s
preferences over all combinations, but ask for preference information only on relevant
combinations. To see this, consider the following example adopted from Sandholm
and Boutlier (2006). For instance, assume that bidder i has indicated she prefers
bundle X over bundle Y, and the lowest cost she is offering for bundle X is 100 €. If the
auctioneer has a better offer for bundle Y from someone else, there is no point for
asking bidder i to express her valuation of bundle Y. This method reduces the number
of packages that need to be valued, but offers no help in the actual valuation process. It
also requires the bidders to submit preference information to the auctioneer.
In some cases firms may prefer not to reveal any preference information to outside
parties as it could reveal the source of their competitive advantage. Therefore,
Hoffman, Menon and van den Heever (2004) have developed a support tool for the
bidders’ private use. The tool is created specifically for the FCC license auctions, but it
could be adjusted for other environments as well. The tool has an interface through
which the bidders insert their preferences in the form of constraints (minimum
population required, overall budget constraint, lowest level of profit acceptable,
minimum bandwidth required etc.). Price information from the current round is then
used to optimize the combinations the bidder should bid on. This second step relies on
the design of the FCC auction, according to which minimum acceptable prices for
each bandwidth are announced after each round.
A somewhat similar, but even more straightforward approach was suggested by Jones
and Koehler (2002). They designed an auction in which bidders only submit rules they
want their bids to follow (i.e. a set of constraints that must be fulfilled). The specific
bids are then constructed by the auction mechanisms when it calculates the optimum
51
allocation. The auction is iterative so the bidders can revise the restrictions they place
on the bids.
Adomavicius and Gupta (2005) offer different kind of bidder support. They do not try
to elicit bidders’ preferences. Instead, they offer metrics by which bidders can assess the
potential of their bids being among the winners. That way the bidders get an idea of
which bids to improve on. However, the metrics do not help in constructing new bids,
because they do not give indication on how to improve the bids.
Teich et al. (2001, 2006) propose a “price support” tool for bidders to use in a multi-
attribute auction, but the tool is also directly applicable in a combinatorial auction.
Based on linear programming (integer programming in combinatorial auctions), the
“suggested price” tool calculates the maximum price for a given combination of items
(or attributes) that brings the bid among the provisional winners. The quantities and
items need to be predetermined by the bidder. Gallien and Wein (2005) present a
similar system and the underlying theory for an optimization-based multi-item auction
mechanism to minimize the buyer’s cost under the suppliers’ (known) capacity
constraints. They assist suppliers in finding a winning bid price. However, the
underlying assumption is that the suppliers are willing to disclose their cost functions to
a supposedly neutral third party auction organizer.
3.2.3 Designing Combinatorial Auction Mechanisms
Mechanism design for combinatorial auctions is not concentrated on the equilibrium
strategies or revenue comparisons between mechanisms, which were the focus of
interest in single-item auctions. The combinatorial environment is so complex that
equilibrium strategies are hard to analyze (Pekeč and Rothkopf, 2003). Some attempts
to construct optimal mechanisms in the Myerson (1981) sense exist, but they are very
limited (de Vries and Vohra, 2003). Instead, combinatorial auction design focuses on
the usability of the auction mechanisms. This involves dealing with the potential
problems identified in the previous section. Also the allocative efficiency of the
mechanisms is considered desirable. The key dilemma in the design is the trade-off
between efficiency and complexity.
52
The elements of mechanism design are the same as in simpler auctions: the way of
communicating bids, determination of winners, and payment rule (see section 1.3).
The first choice is between a one-shot (single-round) and an iterative auction. If an
iterative auction is chosen, it must also be decided whether it is continuous (bids are
allowed to enter at any given time, and WDP is solved after every bid), or round-based
(WDP solved only after each predetermined round). In addition, the designer needs to
decide what information is revealed to the bidders in between the rounds (or bids), that
is, whether bids are open or sealed. The determination of winners is based on revenue
maximization (or cost minimization): the winning bids are the ones that are the most
favorable for the auction owner. There are several possible payment rules, e.g. uniform
pricing, pay-your-bid (first-price), and Vickrey pricing.
The design parameter that affects the mechanism the most is the choice between one-
shot and iterative auction mechanisms. Thus, in the following I will discuss these two
instances separately. Mechanism design literature in combinatorial auctions is
traditionally presented in the forward auction setting, so I will adhere to that. Also, it is
customary to only consider the single-unit case, so that is what I will do as well, unless
mentioned otherwise.
3.2.3.1 One-shot auctions
Multiple-item extensions of single-item, multiple-unit auctions (first-price sealed-bid
auction and uniform-price sealed-bid auction) are not very suitable for a combinatorial
setting, and difficult to implement. In a first-price, sealed-bid auction all combinatorial
bids are submitted before an announced deadline, after which the WDP is solved once
to determine the winners. The winners then pay the amount indicated by their bids.
According to Pekeč and Rothkopf (2003) the benefits of these auctions are that they are
resistant to collusion, and they are transparent as everyone pays the price they bid for.
The main problems are potential computational unmanageability and the complexity
of the bidding task due to strategic complexities and the large number of bids bidders
may wish to construct. Some of the combinatorial auctions implemented in practice,
however, have been one-shot, sealed-bid auctions (e.g. Epstein et al., 2002).
53
When selling multiple units of a single item, market-clearing (uniform) prices have a
certain appeal. However, the idea of market-clearing prices is difficult to translate into
a multiple-item setting (Pekeč and Rothkopf, 2003). The uniform, linear8 market-
clearing prices for the items may not exist, due to the fact that the bidders’ valuations
are superadditive (or subadditive). A simple example adopted from Wurman and
Wellman (2000) illustrates this. Assume that there are two bidders in an auction and
two items for sale. Bidder 1’s valuations are superadditive: she values the individual
items at 0, but has a value of 3 for the pair. Bidder 2’s valuations are subadditive: she
values both items at 2 individually, but gets no extra benefit from obtaining both. The
efficient outcome would be to allocate both items to Bidder 1. However, there are no
linear prices that support this allocation. The prices for the individual items should be
at least 2 in order for Bidder 2 to not be upset over losing, but Bidder 1 would not be
willing to pay 4 for the pair. Due to the superadditive valuations, the linear relaxation
of the WDP may not have an integer solution, and shadow prices for the items do not
exist. In that case, no linear prices exist that would separate winning bids from the
losing ones. Also, in single-item auctions it is customary to use the highest losing bid as
the uniform price, but in a combinatorial auction the concept of a highest losing bid is
not well defined, because bids contain different items.
The Vickrey (second-price) auction has been generalized to combinatorial setting by
Clarke (1971) and Groves (1973). The VCG mechanism is an efficient mechanism
under fairly general conditions (Maasland and Onderstal, 2006), and it can be used in
other frameworks than the combinatorial framework as well. The main restrictions are
that utility must be additively separable in money, and bidders’ valuations must be
independent. In the VCG mechanism bidders announce their valuations over all
bundles (= their type) and the mechanism calculates the optimum allocation and
determines payments.
The payments are determined so that it is a weakly dominating strategy for bidders to
announce their valuations truthfully. This leads to each bidder paying a different price.
The idea behind Vickrey pricing is that bidder i’s payment is the difference in
8 Having linear prices in combinatorial auctions means that the package price is the sum of the item prices in the package.
54
“welfare” of the other bidders without her, and with her in the auction. De Vries and
Vohra (2003) present this formally. Let M denote the set of all items, and S any subset
of M. Thus, vi(S) denotes the value that bidder i attaches to subset S. Additionally, let
y(S, i) = 1 if subset S is allocated to bidder i. Assume that V is the aggregate value from
the auction to the bidders in the optimum allocation y*. Let V-i and y-i denote the
maximum aggregate value and optimum allocation from an auction in which bidder i
is not present. The payment to bidder i is determined by
−−= ∑
⊂
−
MS
iii iSySvVVp ),()( * (3)
where ∑⊂MS
i iSySv ),()( * is the value for bidder i from the winning allocation. Thus, the
term in brackets describes the aggregate value of all the other bidders in the auction in
which bidder i participates in. Another interpretation for the payment is then that it is
the reduction in other bidders’ welfare due to the fact that bidder i by entering the
auction takes a piece of the cake. Notice, that if bidder i is not among the winners (i.e.
y*(S, i) = 0 for all S), V = V-i and her payment is zero. And, if bidder i is the only
winner, her payment equals V-i. Payment is always nonnegative, since V-i (value from
all items M) must be greater than the aggregate value of the subset M\Si (Si = subset
allocated to bidder i) of items to the same set of bidders. A simple example in Pekeč
and Rothkopf (2003) illustrates the VCG payments. Assume that there are two items, a
and b, for sale, and two bidders. The first bidder is offering 10 for {a}, 5 for {b}, and 15
for {a, b}, and the second bidder 1, 6 and 12 respectively. The auctioneer’s revenue is
maximized when item a is sold to the first bidder, and item b to the second bidder
(sum of bids = 16). Without the first bidder the total revenue of the auction would be
12, and her reported valuation for item a is 10, so according to Equation (3) her
payment is 12-[16-10] = 6. Similarly the payment for the second bidder is 15-[16-6] =
5. The total revenue for the auctioneer is 11.
It can be proved that the VCG payments make it a (weakly) dominant strategy (see e.g.
Ausubel and Milgrom, 2006 for a compact proof) to truthfully reveal one’s preferences.
This is the main benefit of the VCG mechanism aside from the fact that it is efficient.
55
It eliminates all gaming elements from the bidding process, so it presumably reduces
the costs of participation.
Even though on a conceptual level the VCG mechanism is very appealing, it has
several disadvantages which make it impractical to implement in practice (Isaac and
James, 2000, Pekeč and Rothkopf, 2003, Ausubel and Milgrom, 2002, Ausubel and
Milgrom, 2006, Maasland and Onderstal, 2006). First of all, the dominant strategy is
far from obvious, especially to inexperienced bidders. In the laboratory experiment of
Isaac and James (2000), only 13.6% of bidders bid their exact valuation, and 49.4% bid
close to their valuation. And even if the bidders knew the dominant strategy, they
might still be unwilling to reveal their valuations to the bid taker. They fear that the bid
taker can use the information in later auctions, and harm the bidders. In larger
auctions, the communication of valuations becomes complicated, as the VCG
mechanism requires bidders to announce their valuation for every conceivable
combination. A valuation should be stated even if the bidder is sure she cannot win a
particular combination, because the payments of the auction depend on losing bids.
Thus, the omission of one losing bid can potentially change the final payments.
Actually, the fact that the final payments are not based on the bidders’ own bids – nor
are they easily identifiable from other bidders’ bids – creates a potential problem. This
is because the bidders may not appreciate the lack of transparency in the pricing. In
fact, the determination of the bidders’ payments requires the solution of an IP problem
for each winner (Porter at al., 2003). The bidders may not trust an auction in which
they cannot verify the mechanism through which their payments were calculated.
Also, even though the Vickrey auction is efficient, it is not necessarily revenue
maximizing. In fact, it can result in low revenues for the auctioneer. This is because
the Vickrey prices are not necessarily in the core of the auction game. A core is the set
of allocations and prices in which the auction owner cannot negotiate a better deal
with the losers (Ausubel and Milgrom, 2002). Consider the following example of
Ausubel and Milgrom (2002). There are two items for sale (A and B), and three bidders
(B1, B2 and B3). B1 only wants the whole package and v1(A,B) = $2 billion. B2 and B3
bid for the single licenses, and v2(A) = v2(B) = v3(A) = v3(B) = $2 billion. In the winning
56
allocation, the items are then allocated to B2 and B3 (one item each). The VCG
payment of both bidders is 2-(4-2) = 0. These prices are not in the core, because the
auction owner would like to go to the losing bidder B1 and offer to sell the items for
her for $2 billion. Bidder B1 would accept this offer, and the revenue for the seller
would increase from 0 to $2 billion. An interesting twist to this example is to consider
what happens if bidder B3 does not enter the auction at all. Now the winner is either
B1 or B2 (they are both tied with a bid of $2 billion), and the price the winner has to
pay is now 2-(2-2) = 2. Such sensitivity of revenue to the number of bidders and the
kind of bids they place is clearly unacceptable. The examples above are of course
extreme examples, but in fact the same phenomena are present whenever the items are
not substitutes for even one of the bidders. Substitutes preferences means that the
bidder’s demand for one item does not decrease when the price of another item
increases. In these examples the items were perfect complements for bidder B1 (a
single item is of no value to her), which violates the substitutes preferences assumption
– with drastic consequences. Recently, researchers have developed auction
mechanisms which would always choose core solutions and prices (Day and Raghavan,
2007, and Day and Milgrom, 2008), but which would choose the VCG prices
whenever they are in the core. This would solve the problem of low revenues of the
VCG mechanism but still preserve the benefits of the VCG mechanism (allocative
efficiency and incentive compatibility).
When the substitutes preferences assumption is violated, the bidders can try to take
advantage of the loopholes making the VCG mechanism susceptible to collusion and
shill bidding by bidders or cheating by the bid taker. Shill bidding refers to the
incentive to use multiple identities or hire someone to pose as a new bidder, and then
buy the items from her after the auction. For instance, assume that there are two items
for sale, and two bidders have made bids of 1000 and 900 for the combination of the
two items. A third bidder, who values the pair at 800, cannot place a competitive bid.
However, if she hires a shill bidder, and they both bid anything above 500 for one of
the goods, they will become winners together. Because there are no other bids for
single items, their payments would be zero. Shill bidding is even easier in internet
auctions, because a bidder can easily enter the auction with multiple identities (Yokoo,
57
Sakurai and Matsubara, 2004). Bidder identification is usually based on information
such as email address, and it is very simple and cheap to acquire multiple email
addresses. The bid taker has an incentive to cheat as well. Once she observes all the
bids in the auction she can increase her revenue by inserting false bids just below the
winning bid prices, and it will be hard for the bidders to detect this.
Even if no cheating occurred, the outcome of the VCG auction might be politically
unacceptable (e.g. if it is an auction collecting revenue for the government). The
public may be outraged when they see that the bidders were willing to pay more (as
indicated by their bid prices), but were in fact charged the VCG prices, which are less
(McMillan, 1994).
Finally, the whole VCG mechanism is designed under the assumption of independent
private values. Common value multiple-item auctions have not been studied
theoretically, but studies of single-item auctions with common value elements (e.g.
Klemperer, 1998) show that Vickrey auctions lead easily to very low revenues to the bid
taker. It is quite reasonable to assume that in many auctions there is either a common
value element to the items or at least one bidder who has complementarities between
the items. Thus, it comes as no real surprise that VCG auctions are not common in
practice.
Rassenti et al. (1982) designed a one-shot, sealed-bid combinatorial auction for
allocating airport time slots. The difference to the Vickrey auction is that Rassenti et al.
use uniform per-item pricing (although bidders announce only package prices). The
uniform prices are more transparent and much simpler to compute than the Vickrey
prices. Because market clearing prices may not exist, it is possible that a package with a
price above the final prices is not accepted causing frustration among the bidders.
Rassenti et al. solve this dilemma by defining two sets of prices: bid rejection prices,
and bid acceptance prices to be announced to the bidders. The acceptance prices
cannot sum up to more than the prices in the winning bids. In case market clearing
prices exist, the two sets converge; otherwise, the bid acceptance prices are lower than
the rejection prices. The abandonment of Vickrey prices means that the auction is no
longer incentive compatible. Rassenti et al. argue, though, that strategic behavior is
58
very risky for the bidders in a sealed-bid, one-shot auction. Nevertheless, it is not easy
for bidders in the auction to determine what kind of bids to place.
3.2.3.2 Iterative Mechanisms9
Iterative mechanisms have distinct advantages over single-round auctions in the
combinatorial setting. The most important advantages are that bidders do not have to
bid for every possible combination in advance, and information can be obtained during
the bidding process (de Vries and Vohra, 2003). The bidders’ task is easier, because
they can place bids when needed, and they can revise them based on feedback
obtained during the auction. Also, if we assume affiliated values, iterative auctions are
more efficient than single-round, sealed-bid auctions due to information revelation
(Parkes, 2006). Most combinatorial mechanisms presented in literature are iterative.
There are already many mechanisms, even though the research field is quite young.
The mechanisms could be classified in many ways, but I will use the classification of
Parkes (2006), who divides iterative mechanisms to price-based and non-price-based
mechanisms. The essential difference between these two groups is that in the price-
based mechanisms bidders are provided with information on what prices to bid for.
Non-Price-Based Mechanisms
Among the first iterative mechanisms is the Adaptive User Selection Mechanism
(AUSM) introduced by Banks, Ledyard and Porter (1989). Bidding in AUSM is
continuous, and bids on all combinations are allowed. The provisional winning bids at
any given time are announced to all bidders. In AUSM, the computational burden is
delegated to the bidders. Anyone willing to submit a new bid must suggest a
combination of bids that complements her own bid, and demonstrate that they
together provide more revenue than the current winning combination. To help
identify good bids, a standby queue is maintained. Bidders can “advertise” their
willingness to make certain bids, and bidders can use these bids as complements to
their own bids. When a new bid comes in, the bid taker has to verify that the
9 Some authors use term iterative auctions as synonymous with round-based auctions (e.g. Kwasnica et al., 2005). However, in this text iterative auctions refer to any kind of auction in which bidders have the opportunity to improve upon their old bids.
59
complementing bids are either in the standby queue or among provisional winners,
and that the new combination in fact produces more revenue. The use of a standby
queue partially alleviates the threshold problem (Kelly and Steinberg, 2000), but the
incentive to try to “free ride” still remains. According to the results of simulations
conducted by Ledyard et al. (1997), auctions using AUSM increased the efficiency of
the final allocations compared to simultaneous and sequential, single-item auctions.
However, AUSM only supports additive-OR bids (Parkes, 2006). This means that
bidders cannot restrict the number of disjoint bids that can become winners. The only
way to indicate e.g. substitutabilities among bids is to make them overlapping. A
version of AUSM was implemented by Sears Logistics to procure trucking services
(Ledyard et al., 2002).
In proxy auctions, automated agents bid on the behalf of the bidders (Ausubel and
Milgrom, 2002). Prior to the auction, bidders express their preferences to the proxy
agents, who then bid to maximize the bidders’ profit. Proxy agents bid until there is no
room for improvement. Winners pay the price bid by the proxy agents (i.e. the lowest
price on a particular combination that allowed the bid to become a winner). Proxy
auctions are efficient provided that bidders can (and are willing to) express their
preferences to the proxy agent. However, communicating preferences to the proxy
agent is every bit as complicated as communicating them to a bid taker in a VCG
auction. Also, no learning can take place in the auction process. The only major
improvement is that they are more resistant to collusion than VCG auctions, and the
failure of the substitutes preferences assumption is not as devastating. In fact, from the
bidders view point, a proxy auction is very similar to a single-round auction, but with
the added option of revising the information given to the proxy agent.
Another problem with the proxy auction is that it can become computationally
infeasible. Whereas in the VCG mechanism, the WDP is solved only once, the proxy
auction advances progressively in increments, and the WDP is solved after every round.
In order to end in an efficient allocation, the increment has to be small enough. This
slows down the convergence of the auction by increasing the number of proxy bids that
have to be placed. The number of rounds could be thousands even for an auction with
six items and ten bidders (Hoffman et al., 2006). Several researchers have suggested
60
methods to speed up the convergence of the proxy auctions (see Hoffman et al., 2006
for a review and comparison).
Another non-price-based mechanism is the direct mechanism suggested by Conen and
Sandholm (2001), where bidders do not have to place bids. Instead, the auctioneer asks
for preference information iteratively from the bidders, but only as little as needed to
determine the optimal allocation.
A clock-proxy auction designed by Ausubel, Cramton and Milgrom (2006) is a hybrid
auction combining a clock auction with a proxy round. In the first stage an ascending
clock auction is organized. The auctioneer announces prices for items, and bidders
report the quantities they demand for those prices. Auctioneer increases prices for
goods with excess demand, and bidders report new quantities. The process continues
until there is no excess demand. The prices established in the clock auction act as a
lower bound on the prices for the proxy round. The clock auction does not allow bids
on combinations, so potential synergies are not realized until in the proxy phase, and
therefore the prices are expected to increase. A clock-proxy auction is actually a hybrid
between price-based and non-price-based auctions. The proxy phase is not price-based,
but bidders receive price information in between the clock auction and the proxy
phase.
Another such hybrid auction with a price-based first stage is the Progressive Adaptive
User Selection Environment (PAUSE) developed by Kelly and Steinberg (2000).
PAUSE is a combination of a simultaneous ascending auction and an AUSM-like
second phase. The simultaneous ascending auction provides information on market
prices, which can then be used in the ensuing AUSM auction, where combinatorial
bidding is allowed. The prices in the package bids in the AUSM phase must be at least
as high as the sum of the prices determined in the simultaneous ascending auction.
Day and Raghavan (2008) on the other hand have designed a three-stage auction, the
purpose of which is to combine the benefits of AUSM and clock-proxy auctions while
avoiding some of their problems (mainly the incentives to free ride and the distortion
caused by linear prices). In the first stage the bidders can submit bid tables in which
they can express substitutabilities between items. In the second stage the can “probe”
61
the auction, that is, ask for prices that would allow specific combinations to become
provisional winners. This is very similar to the “suggested price” tool of Teich et al.
(2001, 2006). The third stage is a proxy auction, after which the winners are
determined. The prices the winners pay are can be either VCG payments or Pareto-
efficient core prices, which ensure incentive compatibility.
Price-Based Mechanisms
As noted earlier, linear market-clearing prices may not exist for a combinatorial
auction. However, different kinds of approximations are available. The approximated
prices have been used in iterative auctions to guide bidders’ bids towards an efficient
allocation. These price approximations can be linear (per-item) approximations, in
which the bundle prices are simply the sums of the per-item prices, or nonlinear
approximations where there are separate approximations for each package. The linear
prices can be used either as per-item “ask prices”, which act as lower bounds on new
bids the bidders create, or “clock prices” at which the bidders announce their most
preferred combinations (= the ones that maximize their payoffs). The non-linear prices
are used only as clock prices. Moreover, the ask prices and clock prices can be either
anonymous, which means that the same prices are announced to all bidders, or non-
anonymous (personalized). In the following I will briefly describe seven combinatorial
mechanisms based on ask prices or clock prices.
Kwasnica et al. (2005) describe a Resource Allocation Design (RAD), which uses
linear, anonymous ask prices. According to the authors, RAD combines in one auction
elements from a simultaneous multi-round auction and AUSM. However, the only
resemblance to these auctions is that RAD allows package bidding (as does AUSM),
and it quotes per-item prices that bidders must beat (as does the simultaneous multi-
round auction of the FCC). Thus, there is only one auction and not two separate stages
as in PAUSE (which also is a combination of a simultaneous auction and AUSM). The
RAD auction proceeds in rounds. After each round, a set of linear prices (one for each
item) is calculated, and new bids must beat these price. Minimum prices for packages
are obtained as sums of the minimum prices for the individual items in the package.
The calculation of the prices are based on three principles: 1) in order to keep “pay-
62
your bid”-feature, the ask prices should be such that the winning bidders end up paying
what they bid for, 2) Prices should be higher than what the losing bidders bid for, 3)
whenever possible the prices should be such that if the losing bidders bid according to
them, they would become winners. When principle 3) holds, ask prices convey
information about opportunities in the auction for the next round, which is desirable.
Kwasnica et al. formulate an LP problem to solve for the prices. The objective of the
problem is to minimize the deviation from the two latter principles (the first principle
must always hold). The prices can be called “pseudo dual prices”, as they are prices
that minimize the deviation from the dual prices of the linearized WDP. One problem
with these pseudo dual prices is that they can oscillate a lot from round to round
(Dunford et al., 2004). This means they can decrease as well, which is counterintuitive.
Thus, Dunford et al. (2004) introduce a smoothed anchoring method to solve for
pseudo dual prices that would deviate as little as possible from the prices quoted in the
previous round.
The Combinatorial (CC) auction of Porter et al. (2003) uses linear prices like RAD.
However, in the CC auction the prices are presented as clock prices. In each round
there is a set of prices at which the bidders are asked to announce their demand for
each item. If there is excess demand for some items, the auction continues to the next
round. Prices for items with excess demand are increased before bidders are asked to
announce their demand. The good aspect about the CC auction is that it is directly
extendable to the multiple-unit setting. However, Porter et al. give a very vague
description on how the final prices are determined in the case when there are no linear
prices to support the winning allocation. Also, Porter et al. (2003) do not tell how the
price increases are determined, so it is impossible to compare the CC auction and
RAD.
Other price-based mechanisms use nonlinear prices, where prices are determined for
each combination and each item separately. Thus the price of a combination does not
have to be the sum of the prices of its components. The benefit of nonlinear prices is
that they do not compromise the efficiency of the final allocation like linear prices do.
However, they are more tedious to solve. Four such mechanisms are AkBA auction by
Wurman and Wellman (2000), iBundle by Parkes (1999) and Parkes and Ungar
63
(2000), dVSV by deVries, Schummer and Vohra (2007), and the Vickrey-Dutch
Auction (VDA) by Mishra and Veeramani (2007).
In the AkBA auction Wurman and Wellman formulate an assignment subproblem that
solves for the minimal (anonymous) prices that support the solution of the WDP.
Supporting prices are defined as prices which, if announced as posted prices for the
combinations, would not cause any changes in the bidders’ behavior. The winners
from the WDP would be willing to purchase the combinations at the posted prices,
and the losers would not. Because there can be a range of such prices, the assignment
subproblem is designed to choose the smallest one to maximize the bidders’ payoff.
The prices are treated as ask prices.
In both iBundle (Parkes, 1999, and Parkes and Ungar, 2000) and dVSV (deVries et al.,
2007) the prices are treated as clock prices, and the bidders are asked to announce their
“demand set”, i.e. the combination(s) which maximize her profit at the current prices.
In addition to being nonlinear, the clock prices in both auction mechanisms are non-
anonymous; that is, each bidder can be announced a different price on the same
combination (although Parkes and Ungar also propose a version of iBundle that uses
anonymous prices). After each round, the prices are only increased for those bidders
and those combinations, which are in the bidder’s demand set, but are not part of the
provisional allocation. In both, iBundle and dVSV, the prices to be increased and the
size of the increase are determined based on an LP solution algorithm. The difference
is that iBundle uses the subgradient algorithm, whereas dVSV uses the primal-dual
algorithm. In both auction mechanisms the winning bidders pay the price indicated in
the bids. However, Mishra and Parkes (2007) develop extensions of both auctions
(extended dVSV and iBEA), in which the winners pay a price lower than in their bid.
The benefit of this minor change is that the new mechanisms lead to efficient
allocations also when the bidders are substitutes condition does not hold.
The Vickrey-Dutch Auction (Mishra and Veeramani, 2007) also uses nonlinear and
non-anonymous clock prices. However, they follow in the Dutch auction logic of
64
decreasing prices in forward auctions, and increasing prices in reverse auctions10. In the
other mechanisms above, prices start low, and increased when there is excess demand.
In the VDA, prices start high, and they are dropped by ε for each bidder for each
combination, which is not in her demand set. The buyer’s demand set contains the
combinations which maximizes her payoff at current prices. The auction ends, when
all the combinations are in the buyers’ demand sets. In essence, the auction reveals the
buyers’ valuations for each combination. The prices in the final price vector are then
adjusted so that they correspond to Vickrey prices. This can be done, because the
auction owner has complete information on the bidders’ valuations. And because the
payments are Vickrey prices, the bidders should not have incentives to misrepresent
their valuations during the auction. Mishra and Veeramani (2007) admit that the VDA
mechanism is not scalable to large auctions, because the number of combinations –
and thereby the size of the price vectors – increases exponentially with the number of
items. However, they do not discuss the fact that bidder may be hesitant to participate
in an auction in which their valuation for all possible combinations – even the ones
they do not win – is revealed to the auction owner. Also, the auction process, in which
announcing your demand for any price facing you gives no indication of whether you
will win or not, may cause frustration among the bidders.
A potential problem with nonlinear clock auctions such as iBundle, dVSV, and VDA is
that clock prices in each round are given only to predetermined bundles. Thus, unless
the auction owner wants to quote prices for each conceivable combination (and have
bidders evaluate all the combinations), the auction can be inefficient. Both
mechanisms lead the auction to an efficient outcome, but only with respect to the
combinations included in the auction. The only way to ensure true efficiency is to
quote prices for each possible combination, which is infeasible in large auctions.
However, even if the auction owner quoted prices for each combination (which is not
possible in multiple-unit combinatorial auctions), the good news for the bidders is that
10 Mishra and Veeeramani (2007) present the VDA mechanism in a reverse setting. Since all the other mechanisms described in this section have been presented in the forward setting, I will transform the mechanism to the forward setting. This should make comparisons to other mechanisms easier.
65
in each round they only have to evaluate the combinations for which the price
changed.
In order to remedy the potential problem, Kwon et al. (2005) propose a mechanism,
the Endogenous Bidding Mechanism, which combines aspects of linear pricing
mechanisms (such as RAD) with iBundle. Their mechanism provides (nonlinear)
prices for combinations, just as in iBundle. In addition to that, they offer a vector of
single-item prices. Bidders can use these single-item prices when constructing new
combinations after the first round. The ask price for any new combination can be
derived from the single-item prices linearly; for the old combinations, the nonlinear ask
prices apply. The mechanism of Kwon et al. (2005) improves the efficiency of iBundle,
because it removes the problem of predetermined bundles discussed in the previous
paragraph. However, the authors are not clear about the performance of their
mechanism relative to RAD.
Table 3 summarizes the key characteristics of the iterative combinatorial auction
mechanisms presented above. I chose to include the one-shot mechanism of Rassenti
et al. (1982) as well, because the way they have calculated the winning and losing
prices resembles that of the iterative mechanisms developed later.
66
Table 3 Summary of combinatorial auction mechanisms
Mechanism Type
Price-
based Type of Prices
Determination
of Payments Bidding
Deriving and
Updating Prices
Rassenti et al.
(1982)one-shot no N/A
linear per-unit prices (sum ≤ bid prices)
package bidding
two sets of prices (for accepted and rejected bids) from the restricted dual
AUSM
Banks et al. (1989)
PAUSE
Kelly and Steinberg
(2000)
RAD
Kwasnica et al. (2005)
Endogenous Bidding
Mechanism
Kwon et al. (2005)
Combinatorial Clock
Auction
Porter et al. (2003)
Ak BA
Wurman and Wellman
(2000)
i Bundle
Parkes and Ungar
(2000)
dVSV
de Vries et al. (2007)
i BEA
Mishra and Parkes
(2007)
DVA
Mishra and Veeramani
(2007)
minimizes infeasibility in the restricted dual
yeslinear, anonymous ask prices
pay-your-bidpackage bidding; bids must satisfy ask prices
iterative (continuous)
iterative (1st stage round-based, 2nd
continuous)
iterative (round-based)
iterative (round-based)
iterative (round-based)
iterative (round-based)
iterative (round-based)
iterative (round-based)
iterative (round-based)
iterative (round-based)
no
1st stage yes, 2nd stage no
yes
yes
yes
yes
yes
yes
yes
N/A
N/A
linear and nonlinear, anonymous ask prices
linear, anonymous clock prices
nonlinear, anonymous ask prices
nonlinear, non-anonymous clock prices
nonlinear, non-anonymous clock prices
nonlinear, non-anonymous clock prices
nonlinear, non-anonymous ask prices
pay-your-bid
pay-your-bid
pay-your-bid
pay-your-bid
pay-your-bid
pay-your-bid
pay-your-bid
discount on bidprice
Vickrey prices
bidders announce profit maximizing packages at current prices
bidders announce profit maximizing packages at current prices
bidders announce profit maximizing packages at current prices
bidders announce profit maximizing packages at current prices
package bidding; bids must satisfy ask prices
bidders announce their demand at current prices
package bidding; bids must satisfy ask prices
minimizes infeasibility in the restricted dual
prices are increased for items with excess demand
minimal prices that support the solution of the WDP
determined through subgradient algorithm
determined through primal-dual algorithm
determined through subgradient algorithm
prices decreased by ε for combinations not in the demand sets
package bidding; bidder must show her bid increases seller revenue
N/A
result of 1st stage is lower limit for 2nd stage
1st stage: bids on single items, 2nd stage: as in AUSM
3.2.4 Combinatorial Auctions in Practice
There are several reports of combinatorial auctions taking place in practice. At least
Sears, Roebuck, and Co. (Ledyard et al., 2002), Mars Inc. (Hohner et al., 2003),
Motorola (Metty et al., 2005) and Procter & Gamble (Sandholm et al., 2006) have
successfully implemented combinatorial auctions in some form in their procurement
67
process. Sears organized combinatorial auctions to acquire transportation services. The
usefulness of combinatorial bidding in transportation auctions has been discussed in
other articles as well (Sheffi, 2004, Caplice and Sheffi, 2006, and Caplice, 2007).
Caplice (2007) reports that since 1997, hundreds of companies have used
combinatorial, electronic auctions to purchase truckload transportation.
Mars Inc., Motorola and Procter & Gamble have used combinatorial auctions to find
suppliers. The reasoning behind adopting auctions as a part of their procurement
process is the same for every firm. The standard practice used to be to negotiate with
each potential supplier individually. These negotiations were lengthy, and it was
difficult for the negotiators to indicate to one supplier what they wanted, since they did
not necessarily know yet, what other suppliers had to offer. Auctions were thought of as
a way to cut down the time and money spent on negotiations, and hopefully to even
find a better set of suppliers to work with. All the firms also had similar concerns about
switching to auctions. They feared that it would ruin the relationships they had with
their suppliers. Indeed, if the reliable suppliers viewed the new auctions as hostile
action attempting to squeeze out all profits, they might choose not to enter the auction.
The auction might not have enough good participants, and the firms would be left
without the required supplies.
The auctions organized by the four firms were all customized to their specific needs,
and therefore they were all different in many ways. However, there were significant
similarities between them too. First of all, they all embraced the complexity of the
environment rather than trying to simplify matters. The auctions were designed so that
they could capture all the benefits from economies of scale and scope the supplier
might have – while making sure that the bidders do not feel being ripped-off. This
leads to complicated designs, in which bidders are able to make very expressive bids on
bundles, but also things like quantity discounts can be expressed. Procter & Gamble
even allow bidders to announce other kinds of conditional discounts. Bidders can also
announce capacity constraints. The buyers can restrict the number of winning bidders
(because dealing with a large number of suppliers is more costly), and they can favor
trusted suppliers, if they wish. All firms report positive results from their preliminary
68
experiences with auctions. Costs have gone down, and the suppliers are still happy
doing business with them.
Combinatorial auctions have also been utilized by the public sector. Epstein et al.
(2002) describe how using combinatorial auctions in procuring meals for schools saved
the Chilean government $40 million annually. More importantly, the reduction in cost
did not come at the expense of the equality of the meals. Another application of
combinatorial auctions in the public sector is the spectrum license auction of the FCC
(Auction #73), which was described already in section 3.1.2.2. Because there has only
been this one combinatorial spectrum license auction so far, it is no possible to
estimate the increase in government revenue resulting from combinatorial bidding.
Auctioning bus routes has become popular in cities and metropolitan areas. Recently,
there have also been bus route auctions allowing combinatorial bidding (Cantillon and
Pesendorfer, 2006, Tukiainen, 2008). Allowing combinatorial bidding makes sense
since there are synergies between bus routes originating at the same place (and possibly
near the garage of the firm). However, both Tukiainen (2008) and Cantillon and
Pesendorfer (2006) report that rather few combinatorial bids were submitted. One
reason for this could be that only a small subset of all bus routes are up for auction
each year, and at least in the case described by Cantillon and Pesendorfer (2006), the
routes were divided into smaller auctions with 4 routes in one auction on average.
69
4 MULTI-ATTRIBUTE AUCTIONS
So far all the auctions I have discussed have been based on price alone (and quantity in
multiple-unit auctions). A completely different take on auctions is to include more
attributes into the bids. Using price as the only bidding attribute is sufficient when
selling existing, clearly defined products, such as agricultural products or works of art.
All other attributes related to the products are predetermined and the information is
available for the bidders. However, often in procurement situations when the item
auctioned does not exist yet and can have many varieties, negotiating merely a price is
not sufficient. The buyer will want to agree on other issues (quality, and terms of
payment and delivery to name a few) before agreeing to sign a contract. Multi-attribute
auctions are designed to take these issues into consideration already in the auction
process. Auctions are often considered a special case of negotiations, and including
multiple attributes into bids brings auctions a step closer to negotiations. Also, as
combinatorial auctions are used in procurement situations, it would be important to be
able to combine these two auction types in one auction mechanism.
Traditionally, companies and governments have sent out requests for quotes (RFQs) in
procurement situations. Based on the quotes, the company chooses the most promising
candidates and begins one-on-one negotiations with them. The contracts are signed as
a result of the negotiations, which can sometimes be long and tedious. Multi-attribute
auctions allow for a more structured process, which should save both the buyer’s and
the seller’s time – and money. In auctions, the rules of the game are clearly defined
and the bidders are aware of the attributes according to which their bids are evaluated.
The process becomes more automated in a way, especially if the Internet is used as the
medium for the auction. Multi-attribute auctions are flexible in the sense that they can
be single-unit, multiple-unit11 or even multiple-item auctions.
Multi-attribute auctions resemble combinatorial auctions in some ways. First of all,
bids in both auctions are vectors. Only now the vector components indicate the levels
of attributes, where as in combinatorial auctions they indicated desired quantities of
11 Teich et al. (2004) use the term “multiple issue auction” when referring to a multi-unit, multi-attribute auction.
70
different items. Secondly, winner determination is complicated in both auctions,
although for different reasons. In combinatorial auctions the difficulties were
computational. In multi-attribute auctions the difficulty is in comparing the bids with
each other – it is like trying to compare apples and oranges. The bid taker’s preferences
over the multiple attributes need to be elicited prior to the auction, and the
mechanism designed according to these preferences. Again there is a certain
resemblance to combinatorial auctions, where the bidders needed to elicit their
preferences over the different items in the auction. The problem is essentially the
same, even though this time it is the bid taker, who has to solve it. Naturally, the
bidders also need to make similar evaluations when placing bids. The fields of multiple
criteria decision making and decision support specialize in developing tools to help
decision makers express their preferences over multiple attributes. In the following
sections I will review different approaches to multi-attribute auction design found in
auction literature. The review will follow along the lines of Teich, Wallenius,
Wallenius and Koppius (2004). First I will present the scoring function method, which
is based on value function theory and commonly found in theoretical articles. Then I
will proceed to less rigorous and perhaps more user-friendly auction designs.
4.1 The Scoring Function Approach
One way to evaluate multi-attribute bids is to formulate a scoring function S(x): Rn →
R, where x is the vector of n attributes. The scoring function assigns each bid a score
based on which the bids can then be ranked. The winner of the auction is simply the
bidder whose bid produces the highest score. The concept of the scoring function is
similar to that of the value function. The construction of value functions has been
extensively studied within the multi-attribute utility theory (MAUT). The value
function assigns a value for the level of each attribute and then combines the values
from all attributes. A commonly used value function is the additive value function with
weights.
The focus of auction studies implementing the scoring function differs from that of the
MAUT. Where the MAUT studies different approaches to the elicitation of preferences
and construction of value functions, most theoretical auction studies take the scoring
71
functions as given. The point of interest in auction papers is either the optimal (i.e.
utility maximizing) scoring rule under different auction mechanisms (Che, 1993,
Branco, 1997, and Beil and Wein, 2003) or the study of the economic implications of
the multi-attribute setting compared to the single-attribute auctions (Bichler, 2000, and
Chen-Ritzo, Harrison, Kwasnica and Thomas, 2005).
The scoring function approach follows the research tradition established by the game-
theoretical price-only auction studies. The scoring function reduces the multi-attribute
auction to a single attribute auction, which allows the development of mathematically
beautiful models similar to auction models such as the IPV and Common Value
models. In other words, the scoring function auction models are an extension of the
traditional models. This is apparent in the objectives of the studies, which include
testing the efficiency of different mechanisms and the revenue generated by the
auctions, and in the formulations of the game-theoretic models.
Bichler’s (2000) approach is more practice-oriented even though the objectives of his
study are related to the efficiency of the outcome and the comparison of the payoffs of
different auction mechanisms. The major difference is that Bichler has not only
designed an auction, but also implemented it. He tests the payoff equivalence
hypothesis and the efficiency of a multi-attribute auction (using a scoring function) in
the WWW-environment using MBA students as test cases. Bichler arrives at the
conclusion that multi-attribute auctions produced higher payoffs than single-attribute
auctions. However, he does not mention, how the values for the non-price attributes
were chosen in the single-attribute auctions. Hence the basis of such a comparison is
left vague. More convincing are Bichler’s results indicating that the auction
mechanisms were not payoff equivalent (first-score auction produced a higher revenue
than English or second-score auctions). Bichler offers the heterogeneity (asymmetry) of
bidders as a possible explanation for the discrepancy between theory and the results of
the experiment.
Following Bichler’s example, Chen-Ritzo et al. (2005) also compare multi-attribute
auctions to price-only auctions. Their experiment is more complicated than Bichler’s
(they use three attributes instead of two), but their results are the same: multi-attribute
72
auctions are more efficient and result in higher payoffs. Chen-Ritzo et al. (2005) use
the utility maximizing levels of the non-price attributes in the price-only auctions,
which makes their comparisons more reliable than Bichler’s (2000).
4.2 Other Approaches
The explicit assessment of value functions, which the scoring function approach
requires, has been criticized by several researchers over the years (e.g. Simon 1955,
Larichev, 1984, and Korhonen and Wallenius, 1996). It requires practice and expertise
to be able to express one’s preferences in the form of a value function. In the auction
setting, the assessment of the auction owner’s preferences should be easy in order to
entice managers to resort to auctions in the procurement process. The applicability of
the scoring function method is thus questionable. Bichler (2000) added a decision-
aiding tool to help auction owners construct their value functions. However, he does
not describe the tool. If the process is not transparent and understandable to the user, it
will not evoke trust in the procurement manager and she might decide not to use the
auction system.
The focus of recent studies in multi-attribute auctions has diverted from the focus of
traditional auction research. Now the focus is not as much the efficiency or the
revenue (utility) equivalency of auction mechanisms, but the functionality of the
designed auction. Functionality refers here to the ease of use for both the seller and the
buyers. Even though the primary goal is not to design an auction that maximizes the
auctioneer’s revenue (utility), it is, of course, important that the auctioneer receives a
satisfactory utility. Multi-attribute auctions can be seen as cooperative negotiations (see
e.g. Guttman and Maes, 1998), as there can be opportunities for joint gains. In a way,
multi-attribute auctions resemble traditional one-on-one negotiations. Therefore, some
multi-attribute auction design builds upon negotiation theory, e.g. the The Leap Frog
Method and the Auction Owner Controlled Bid Mechanism suggested by Teich,
Wallenius and Wallenius (1999).
Cripps and Ireland (1994) propose a method, which uses quality thresholds. This
would eliminate all the other attributes besides price, and render the auction to the
price-only situation. They consider three specific designs. In (a) a price-only auction is
73
held only after bidders have submitted their quality plans (which contain information
on all non-price attributes) and the plans have been accepted. In (b) the auction is held
first, and quality plans are requested in the order indicated by the auction results,
starting with the winner. The first bidder, whose quality plan is approved, obtains the
contract. In (c) price and quality plans are submitted simultaneously, and the contract
is awarded to the best priced (i.e. the cheapest deal in the reverse auction setting) plan
that satisfies the predetermined quality requirements.
Teich, Wallenius, Wallenius and Zaitsev (2001 and 2006) have implemented the
“pricing out” method in their Internet-based hybrid auction called NegotiAuctionTM.
Pricing out can be used without having to explicitly formulate the auctioneer’s value
function. Instead, it probes into the implicit preferences of decision makers. This
makes it a popular approach among decision analysts. It is, however, possible to
construct a value function with pricing out, but Teich et al. (2001, 2006) choose not to
do so.
Pricing out can be used in all situations, where there is a natural monetary attribute
(price or cost) attached to the auctioned asset. Borcherding, Eppel, and von
Winterfeldt (1991) compared pricing out with three other value elicitation techniques.
They concluded that weights generated with pricing out corresponded closest with
weights generated externally by a group of experts.
Keeney and Raiffa (1976) provide a general description of the pricing out method. The
underlying idea is to express the decision maker’s (in this case the auction owner’s)
preferences over multiple attributes in monetary terms. Assume that there is a
monetary attribute M and n different non-monetary attributes X1, X2, … , Xn related to a
product. Lower case letters m and x1, … , xn denote the values the attributes can
assume. In pricing out the auction owner is asked to identify the monetary value m* for
a bundle x*=(x1
*, … , xn
*) which makes her indifferent between the bundle (m*, x*) and
a predetermined reference bundle (m0, x0). That is, (m*, x*) ~ (m0, x0). The difference m*
- m0 depicts the auctioneer’s willingness to pay for the possibility of transforming the
bundle x0 into x*.
74
The approach explained above becomes tedious and time-consuming when the
number of vectors that need to be evaluated is large. Then, it pays off to make
simplifying assumptions. Keeney and Raiffa (1976) state that pricing out can be made
easier when
1) The difference between m* and m0 (i.e. the willingness to pay) does not functionally depend on the base value of m0
2) The monetary attribute M and attribute Xi as a pair are preferentially independent of the complementary set of attributes
When these assumptions apply, the pricing out can be done individually for each
attribute. The method is the same as in the general case described above, but here the
value of only one attribute will be changed at a time. For each attribute Xi the
auctioneer is asked to state a monetary value m* that will satisfy the indifference
equation
(m*, x1
0, … , xi-1
0, xi
*, xi+1
0, … , xn
0) ~ (m0, x0) (4)
Thus the auctioneer only has to make n such assessments instead of pricing out all
possible combinations of the attributes X. This simplifies the preference elicitation
process and formulates it in such a way that it is easy for the auctioneer to make the
assessments.
Pricing out also provides computational advantages, because it reduces the bids to two-
dimensional vectors containing only price and quantity components. The bidders still
submit multi-dimensional bids, but all other attributes are “priced out” before the
winner determination problem is solved.
The various preference elicitation methods described above try to achieve two goals:
the realistic and truthful description of the preferences of buyers and sellers and the
ease of use of the method for all participants. Unfortunately, most of the time these
goals conflict. The more elaborate the preference elicitation scheme is, the more
difficult it is for a novice to use. The scoring method clearly emphasizes the first goal
and is therefore more suitable for theoretical studies. The rest of the above mentioned
approaches prioritize the usability of the methods attempting to generate mechanisms
that could be implemented in practice. Regardless of the method in question, it is clear
75
that in the multiple-issue case, bidding becomes more difficult for the bidders,
especially inexperienced ones. Also, it is more difficult for the auction owner to set up
an auction that would produce the kind of results that would match her true
preferences. Hence, all sorts of decision aid tools become important in auctions
(Bichler, 2000), and decision making theory becomes relevant for auction theory. The
introduction of Internet auctions provides an excellent medium to include decision
support with auctions, as will be discussed later in Chapter 5.
4.3 Multiple Attributes in Combinatorial Auctions
Since both multi-attribute auctions and combinatorial auctions are very suitable for
procurement situations, it would make sense to combine aspects of both auction types
into one auction. However, as both auction types alone are already quite complicated,
their combination cannot be expected to be any simpler. There has not been much
work on multi-attribute combinatorial auctions, but researchers have identified the
potential benefits of such combinatorial auctions.
Sandholm and Suri (2006) propose the use of a weighting function to translate the
multiple attributes into monetary terms. The weighting function f(pj, aj) takes a bid bj
and weights its price pj with the values of the attribute vector aj and returns a new price,
which is then used to compare bids with each other. The approach suggested by
Sandholm and Suri bears great resemblance to the pricing out method used by Teich
et al. (2001, 2006). Teich et al. present the pricing out method in a single-item setting,
but the extension to a multi-item auction is straightforward. In a multi-attribute
combinatorial auction the incoming bid would actually be a matrix, where the
columns indicate the items, the last column being the price, and the rows the values of
different attributes (first row is for the item quantities and subsequent ones for non-
price attributes).
Epstein et al. (2002) describe a practical application of a multi-attribute combinatorial
auction. In the auction the Chilean government procured school meals. The auction
was designed to be as simple as possible in such a complex setting. Thus, the multiple
attributes were taken care of by pre-auction approval. The government set several
criteria for the food to be delivered, and only firms able to fulfill the criteria were
76
allowed to place bids. This approach is similar to the quality threshold method
proposed by Cripps and Ireland (1994).
The “expressive bidding” procedure of Sandholm (2007) is an ambitious attempt to
combine combinatorial and multi-attribute auctions for procurement purposes. His
auction system, CombineNet, also allows the bidders and the auction owner to include
a variety of side constraints (capacity constraints, minimum or maximum number of
winning bidders, etc.). The system also allows the bidders to express discount
schedules. The underlying goal in the development of CombineNet is create as much
flexibility as possible so that the bidders and the buyer could find win-win solutions
(Pareto improvements).
77
5 ONLINE AUCTIONS AND AUCTIONS IN PRACTICE
A big part of auction literature concentrates on auctions in theory, and forget about the
practice. According to Klemperer (2002) most of the traditional studies (also reviewed
in this thesis) are of little help when designing auctions in practice.
The development of computers and the introduction of the Internet have had a
tremendous impact on the practical side of organizing auctions, and thereby also on
auction research. First of all, computers have enabled the organization of new kinds of
auctions, like combinatorial and multi-attribute auctions, and they are helpful in
multiple-item auctions (Pinker et al., 2003). Secondly, besides enabling the use of
more complicated auction mechanisms, the Internet has affected the traditional and
simple single-item auctions creating new potential problems for the design.
The Internet has introduced new elements into auctions, which were not present in
traditional auction settings. First of all, Internet has reduced transaction costs from
organizing auctions and participating in them. This has broadened the spectrum of
products that can be sold through auctions. Many standard products, which earlier
were always sold with posted prices, are nowadays also sold through auctions. Because
the items sold in online auctions are not unique, it is likely that similar products are
sold in several separate auctions. Some of these auctions can be ongoing at the same
time, and some occur later in time. In any case, it is not reasonable to treat such
auctions as isolated and independent incidents.
Other new elements include an increased number of potential participants, and the
endogenous entry of new bidders (i.e. bidders can enter auctions even after they have
begun). Also, the most common auction type in the online environment seems to be a
uniform price multiple-unit auction (Bapna, Goes and Gupta, 2000). This is in clear
contrast to the pre-Internet era, when the single-unit auction was the prevalent auction
type. Also, the duration of online auctions can be much longer than that of traditional
auctions. All of the new elements mentioned above affect the analysis of all auctions,
and the theoretical results of traditional auction theory focused on single-unit auctions
(e.g. Vickrey, 1961, Myerson, 1981, and Milgrom and Weber, 1982) may not apply
78
anymore. According to Pinker et al. (2003), there has been very little research so far
that would consider these new elements.
Auctions, even the more complex ones, allegedly have lower transaction costs than
negotiations. There are some concerns, though (Pinker et al., 2003). First of all,
procurement auctions have been accused of squeezing out all surplus from the bidders.
The second concern is that people advocating the low transaction costs do not take into
consideration all costs related to procurement by auction. Mainly these ignored costs
refer to switching costs incurred whenever changing to a new supplier. Also, time costs
from participation can be significant. The third argument is that the trend in supply
chain management has favored vertical integration and partnerships, and auctions do
not support this development. Tight partnerships cannot be established, if the supplier
base is renewed annually based on results of auctions. All of these concerns were
voiced in the cases of Mars, Inc, Motorola and Procter & Gamble (see section 3.2.4).
The potential problems were considered when designing the auction, and the results
have been positive on all accounts.
The fact that auctions are used more often in different kinds of market transactions,
and the demand for different kinds of auction designs has grown, there is also need for
more research on auction design in practice. The fact is that designing a successful
auction is by no means trivial. Already the choice of the appropriate type (single- or
multiple-item, single- or multiple-unit, multi-attribute or price-only) and mechanism
(English, Dutch, first-price sealed-bid, Vickrey, and their extensions) is difficult, and
abundant theoretical research is not necessarily of much help, because it relies on a set
of restrictive assumptions. Once the auction type and mechanism are chosen, they
already determine some of the major design issues such as bid type, pricing rule and
winner determination. However, this is not sufficient. One must also take into
consideration minor issues, which are not inbuilt in the auction mechanism. In the
following I will discuss additional design issues that have not been presented yet,
because they have not been considered in the theoretical models. I will also discuss the
need for additional rules for the auction, the purpose of which is to minimize the risk
for the auctioneer from bidders cheating or defaulting.
79
5.1 Additional Design Issues
Additional design issues are issues that need to be determined when designing an
auction, but which are not tied to any particular auction mechanism. Such issues are
the bid increment (or decrement in descending price auctions), possible reservation
price and the complete information architecture of the auction.
The bid increment is the minimum amount by which the new bid must exceed the
current highest bid in order to become the new provisional winner. Similarly, the bid
decrement is the minimum amount by which the price must decrease (in a reverse
auction) before the bid can become a provisional winner. In theoretical models the
bids are assumed continuous, that is, an increase of the size of ε is always possible, but
in practice in almost every auction a fixed increment is defined. The increment can be
of either a fixed dollar value or a percentage. In high value auctions, an increment of
1% could be millions of dollars, hence there usually is a maximum dollar value for the
increment as well. The size of the increment/decrement naturally affects the
convergence speed of the auction. The larger the increment, the fewer bids are needed
to reach the final outcome. However, the larger the increment the larger the expected
gap between the auctioneer’s actual revenue and potential revenue. For example,
assume that the bidder with the highest valuation has the valuation of 100, and the bid
decrement is 10. If the current highest bid in the auction is 91, the bidder cannot bid
anymore. If the increment were 9 or less, the bidder could place a bid, and the
auctioneer’s revenue would increase. The choice of bid increment/decrement has not
been studied much. Teich et al. (2001, 2006) use a bid decrement in their formulation
of the NegotiAuction system, but they do not study the effect of different decrements.
Bapna, Goes and Gupta (2000, 2003) also have a bid increment in their auction
mechanism. Bapna et al. (2003) conclude based on theoretical and empirical
considerations that the bid increment has a significant effect the auctioneer’s revenue,
but their result is tied to the multiple-unit, uniform-price auction. Also, the bid
increment in the studied actual auctions is devised as a fixed dollar amount and not
proportional to the bid amount. However, clearly the role of the bid increment should
be studied more.
80
The reservation price means the price below which the auctioneer will not sell the
item (in a forward auction) or the price above which the auctioneer will not buy the
item (in a reverse auction). The reservation price was mentioned already in the context
of optimal auctions in section 2.5. In the case of optimal auctions the reservation price
is set to equal the expected valuation of the second highest bidder. However, such a
decision rule is of little help, if the distribution of the bidders’ valuations in unknown –
provided that there even exists such a distribution. The main idea behind setting a
reservation price is to shield the auctioneer from an undesirable outcome. For
example, little competition can lead to very low prices in a forward auction. A good
example of this is the spectrum auction in New Zealand (see section 2.4) in which the
winner paid a very low price because the difference between the highest and second
highest bids was huge. A reservation price anywhere in the gap would have guaranteed
the government a higher revenue. The downside of the reservation price is that it can
discourage bidders from entering, and the item might not get sold at all. An additional
question related to the reservation price is whether to disclose it to the bidders or keep
it as a secret. Bajari and Hortaçsu (2003) have discovered that in eBay auctions the
number of participants decreases, revenues are lower, and items are left unsold more
often, if a secret reservation price is applied.
In online auctions, the duration of the auction also becomes a design issue. In
traditional auctions, which take place in auction houses, the duration of the auction is
measured in minutes, and there is no need to predetermine the duration. However, in
online auctions where the bidders are not present when the auction opens, the auction
duration becomes an issue. Potential bidders need to have time to find the auction.
The longer the duration, the more bidders find the auction and participate in it (Vakrat
and Seidmann, 2000). Hence, in principle the number of participants should increase
as the duration increases, which should be better for the auction owner. However, if
the duration is very long, the bidders do not have an incentive to bid at the beginning.
By bidding early, they only run the risk of increasing the price. Bidders prefer waiting
until the auction is about to close to observe the level of competition. Thus, it is
possible that the number of bids remains low. Also, the longer the auction, the larger
the transaction costs for the participants. Bidders need to monitor the auction, plan
81
their strategies and place bids. In high stakes auctions firms usually have teams of
experts involved in the bidding process throughout the auction, which can be a big cost
to the firms. Take for example the FCC auctions in which all major telecom
companies had hired their own team of auction researchers to advise them in the
auctions.
Information architecture as understood by Koppius and van Heck (2002) encompasses
all information flowing between the bidders and the auctioneer. Part of the
information architecture is thus determined by the auction mechanism, for example
the type of bids (multiple-item, multi-attribute or price-only) and whether the bids are
openly visible to every participant (open-cry auctions) or kept a secret (sealed-bid
auctions). There can be various degrees of bid openness, though. In some auctions
only the provisionally winning bids are announced and in others the bidders only know
the status of their own bids (provisional winner or not), but not the content of any other
bids. Information architecture also contains all the information the bid taker wishes to
disclose to the bidders. This can be the number and identity of all the bidders, the bid
taker’s reservation price, or the bid taker’s preferences over the multiple attributes in
multi-attribute auctions. The bid taker can also choose to misrepresent some
information, or reveal only partial information about her preferences. Koppius and van
Heck study the impact of information architecture in multi-attribute auctions.
According to them, information architecture is even more important in multi-attribute
auctions, because it enables the bidders and bid taker to identify possibilities for Pareto-
improvements. Thus, they hypothesize that more information would improve the
efficiency of the final allocation. Their experiments support this hypothesis, but they
also discover a saturation point after which more information did not affect the auction
outcome significantly.
5.2 Auction Rules
When organizing a real auction, choosing the mechanism and designing the details is
not enough; some fine tuning is required. The need for more detailed rules is
especially pronounced in electronic auctions, because it becomes more difficult to
observe and control bidder behavior and prevent cheating. In this section, I will discuss
82
rules, which concern issues and behavior that are ignored in theoretical models (e.g.
defaulting on bids, cheating and collusion), but which are important in practical
applications. There has been a lot of discussion on auction rules in the context of the
FCC Auction #31 (FCC, 2000, Pekeč and Rothkopf, 2000, Vohra and Weber, 2000),
which ironically after all the planning and discussions never took place. Also articles by
Pekeč and Rothkopf (2003), Kelly and Steinberg (2000) discuss different auction rules,
and Klemperer (2002) describes several real auctions where design flaws have led to
undesirable auction outcomes.
The reasoning behind rule design is that bidders will try to take advantage of the
auction design in any way they can think of. For instance, in iterative auctions with a
predetermined ending time bidders may want to wait until the last minute before
making a bid. This type of behavior is known as sniping. Snipers hope to be able to
surprise the competition with a last minute bid and leave no time for the competitors
to respond (Roth and Ockenfels, 2002). If they are successful, the revenue from the
auction remains low. To combat this kind of behavior, activity rules have been
introduced in iterative auctions which involve substantial amounts of money. Activity
rules state how often and what kind of bids a bidder must enter in order to be allowed
to continue bidding. Activity rules are often linked to minimum bid increments.
Closing rules affect activity rules. Activity rules become more important, if the auction
has a predetermined closing time. However, if the auction closes only after a certain
period of inactivity, activity rules are not as critical, but it does not mean they would be
unimportant. In an extreme case there would be no bidding activity until right before
the closing time. There would still be competition as the ending would be pushed
back, but the process of price discovery would not be as efficient undermining one of
he benefits of iterative mechanisms. Auction owners may want to use eligibility rules as
well. Eligibility rules require the bidders to prove their solvency prior to entering the
auction. For instance, they might have to make a deposit, and the size of the deposit
determines the size of bids they are allowed to place.
Collusion in the form of collusion rings agreed upon prior to the auction was discussed
already in section 2.4. Collusion does not have to be explicit, though. During the
auction bidders can try to signal their intentions to competitors. Maasland and
83
Onderstal (2006) and Klemperer (2002) report a case of signaling in the German GSM
auction. In 1999, the German government put 10 licenses up for auction, and required
a 10% bid increment. One of the competing firms, Mannesmann, made a bid of 20
million DM on five licenses, and a bid of 18,8 million DM on the other five licenses.
Mannesmann’s main competitor T-Mobile was able to calculate that topping the
current high bid of 18,8 million with the required bid increment of 10% would bring
the price of these licenses to about 20 million as well. Actually, this was Mannesmann’s
way of signaling to T-Mobile its willingness to share the licenses equally; it did not
necessarily have to bid for the other five licenses at all. T-Mobile understood the signal,
and the auction ended very shortly with the bidders sharing the licenses at the price of
20 million DM. The efficiency of the outcome is hard to determine, but clearly the
final price obtained by the seller (German government) was artificially low. This
example shows that no explicit collusion among the bidders is necessarily needed for
them to reach a silent agreement to not drive up the prices.
In some auctions bidders have used the lower digits of bid prices to signal the items
they are interested in (Kelly and Steinberg, 2000). To combat this, the auction owner
can require that bid prices follow certain increments. Of course, sealed-bid auctions do
not suffer from this type of signaling. Another form of signaling is called jump bidding,
which was observed in earlier FCC auctions (McAfee and McMillan, 1996). Jump
bidding refers to aggressive bidding behavior, where a bidder places a bid way above
the required increment in order to warn other bidders. One way to prevent jump
bidding is to use a clock auction, in which the bidders simply announce their demand
at the price indicated by the auction clock instead of calling out their own prices
(Banks et al., 2003).
Cheating and fraud are also a concern in auctions, and especially in online auctions. It
is easier for bidders to remain anonymous, and thereby use multiple bidder identities
in the same auction (Pinker et al., 2003, Yokoo et al., 2004). In a similar vein, it is also
possible for the bid taker to place bids under a false identity in order to force the
auction to a more favorable outcome. These cases of cheating were discussed already
in the context of the VCG mechanism in section 3.2.3.1. Bidders could also choose
not to pay, or sellers refuse to provide the product after receiving a payment. Auction
84
designers have tried to design mechanisms to detect and prevent fraud. Also reputation
of both buyers and sellers has become a factor in the electronic market places.
Pekeč and Rothkopf (2003) point out that auction rules should also define how ties are
to be broken. Theoretically ties are not interesting, because with continuous
distributions their probability is zero. In practice, however, it is quite possible that a tie
arises, because bidders tend to round prices, or the auctioneer requires bidding at
certain bid increments (decrements). In combinatorial auctions it is even more likely
because ties can occur in different ways, because the same total revenue can result
from many different combinations. In the name of fairness, the ties should be broken
based on predetermined rules. One way to break ties is to pick the winning
combination randomly. A more sophisticated way is to use time stamps. Each
incoming bid receives a time stamp when it enters the auction. The winning
combination is then either the one that was completed first (highest time stamp value
is lower than the highest stamp value of other combinations), or the one with the
lowest average time stamp. Using the time stamps may not be entirely fair though,
because due to differences in traffic loads on the Internet, some bids might be at a
disadvantage.
Additional issues to be decided on are how to deal with bid withdrawals or defaulting
on winning bids. Withdrawals are usually allowed in simultaneous and sequential
auctions to alleviate the exposure problem. However, in combinatorial auctions
exposure problem is not as crucial, so allowing withdrawals may only make collusion or
signaling easier. Penalizing withdrawals and defaulting on winning bids can thus help
reduce cheating and “gaming” behavior. Setting eligibility rules and requiring the
bidders to make a deposit also ensures that the bidders are capable of paying the
penalties.
The important thing to keep in mind in auction design is that each auction is unique
and therefore requires a unique combination of design parameters. This fact is
emphasized by Binmore and Klemperer (2002), who report their experiences on
telecom license auctions in the UK and other European countries. According to them
it is not enough that the items for sale are identical to justify using an identical auction
85
design, because market conditions (number of potential participants, attractiveness of
market) vary. Sometimes it is enough to adjust minor rules, but in some cases the
whole auction mechanisms needs to be modified. In Binmore and Klemperer’s study,
what worked well in the British telecom auctions, did not work in the Netherlands or
Switzerland, which clearly demonstrated that auction design should not be copied.
86
III MOTIVATION AND OBJECTIVES
6 MOTIVATION AND OBJECTIVES
As discussed in the literature review, combinatorial auctions are characterized by
complexity. The winner determination is computationally complex, and the
construction of bids requires the elicitation of complex preferences over a set of
different items. Good bidding strategies are not easy to calculate, as good bids for one
bidder depend not only on her preferences and cost structure, but also on bids made by
other bidders. There is a lot of literature concerning the efficient solution of the WDP
(see section 3.2.2.1), and recently there has also been some research into the elicitation
of bidder preferences (section 3.2.2.2) and the design of auction mechanisms (section
3.2.3). The addition of multiple units of each item makes the mechanism design more
complicated – and complicates further the bidders’ task of submitting bids. Also, only a
few of the mechanisms are easily extended to the multi-unit setting. Bidder support has
been neglected in existing literature. Thus, there is a need for a multi-unit auction
combinatorial, which would be easy for bidders to participate in.
6.1 Need for Mechanisms for Multiple-Unit Combinatorial Auctions
Thus far, the extension of the combinatorial auction mechanisms for multiple-unit
cases has not been discussed in literature much. Out of the mechanisms reviewed in
section 3.2.3.2 five are easily extendable to the multi-unit setting: AUSM, PAUSE,
RAD, the Endogenous Bidding Mechanism, and the combinatorial clock auction.
However, all of them have their shortcomings, one of them being that they
compromise efficiency. In multiple unit auctions, the effect of linear ask prices on
efficiency would be even worse because they remove the possibility of expressing
economies of scale (i.e. quantity discounts). Using non-linear pricing improves
efficiency, but it cannot really be used in multi-unit auctions, or large single-unit
auctions for that matter. This is because auctions with multiple units, or a large
number of items have too many possible combinations to be evaluated in reasonable
time. Of course, not all combinations need to be considered – e.g. iBundle only
considers combinations for which there are bids from previous rounds – but that
87
compromises efficiency. Thus, there is room for a different approach in combinatorial
auction mechanism design.
6.2 Identification of the Puzzle Problem and Need for Quantity
Support
Combinatorial auction literature recognizes the threshold problem – as noted in
section 3.2.1.3 – but little has been done to try to alleviate it. In addition, I believe the
threshold problem does not adequately describe all the problems facing the bidders
bidding in a combinatorial auction. Firstly, the threshold problem refers to a situation
in which a large number of “local” bidders – bidders bidding for single items or small
packages – are trying to coordinate their bid prices to outbid a “global” bidder – a
bidder bidding for the whole bundle (or a few big bidders). However, the bid price is
only one dimension in the bid vector in a combinatorial auction. The bidder can also
choose to vary the item combination associated with the price, and this, I believe,
creates a whole new problem.
In combinatorial auctions, a successful bid complements existing bids, placing all of
them among the winners (unless the winning bid is for the entire bundle). This brings
a cooperative flavor into the auction even though bidders are still in competition with
each other and are not allowed to collude. The threshold problem is one phenomenon
arising from this cooperative nature of bidding. The threshold problem – the way it is
presented in literature – is confined to price adjustments. However, in combinatorial
auctions, the item combination in a bid plays as large a role in determining whether
the bid is among the winners or not. A combinatorial auction is like a puzzle: in
addition to the prices being right, the bids need to fit together to form the whole
bundle like puzzle pieces fit together to form a complete puzzle. However, in a
combinatorial auction, the size and shape of the puzzle pieces are not predetermined;
it is the task of the auction mechanism to endogenously determine them. The WDP is
then analogous to the process of choosing which pieces to use to compile the puzzle.
Due to this very fitting analogy, I call the problem of finding and placing bids that
complement other bidders’ bids (i.e. bids that will be chosen by the WDP), the puzzle
problem. An implication of the puzzle problem is that even if a bidder has managed to
88
identify her most preferred combinations, it may not make sense to bid on them, if
there are no complementing bids coming from other bidders.
In open-cry auctions, the puzzle problem is not more difficult to overcome than the
threshold problem. But in sealed-bid auctions the puzzle problem becomes almost
impossible to overcome, and it is a serious threat to allocative efficiency in such
auctions. In sealed-bid auctions the bidders do not know the contents of the competing
bids, and thus it is impossible for them to deliberately place complementing bids. The
puzzle problem can thus arise without the coordination issues, which are at the heart
of the threshold problem. A bidder could be able to place a bid that would make her
(and a group of existing bids) winners, but does not know which bid it would be.
Most procurement auctions use a sealed-bid format, because they are preferred by the
bidders (Jap, 2003). In sealed-bid auctions bidders do not have to worry what
information their bids could reveal to their competitors. Also, for the auction owner a
sealed-bid auction has the advantage that it removes the possibility of signaling, and
jump bidding will not be as effective because competitors cannot observe it. The fact
that combinatorial procurement auctions are often held in a sealed-bid format means
that the concerns of auction outcomes being inefficient due to the puzzle problem are
very relevant. The problems are aggravated in multiple-unit combinatorial auctions,
because not only do the items in the bids complement each other, but the quantities
also need to add up to the total demand. Thus, it can easily happen that a bidder with a
low cost structure (in a reverse auction) loses, because she did not bid for the “right”
combination.
Even though winning in a combinatorial auction depends on other bidders’ bids, the
support mechanisms presented in section 3.2.2.2 or the iterative auction mechanisms
in section 3.2.3.2 do not attempt to find bids to form coalitions with other bidders.
Offering price information in an iterative auction guides the bidders to bid for items
with a relatively high price, but it does not help in determining the quantities for each
item (this is of course relevant only in multi-unit cases). I feel that this aspect of
bidding in combinatorial auctions has been neglected in existing literature, and it is
important that the puzzle problem be addressed.
89
6.3 Objectives and Methods of the Study
The objective of this study is to overcome the puzzle problem present in multi-unit
sealed-bid and semi-sealed-bid12 combinatorial auctions. In this research project, we
consider only iterative auctions, as one-shot auctions present very limited opportunities
to a) gather any kind of information from the auction, and b) to support bidders. Also,
our focus is on continuous, iterative auctions, and not on round-based auctions.
We try to reach the objective by developing support tools for bidders. The task of the
support tools is to find bid suggestions that would complement the existing bids (that is,
identify the size and shape of possible missing pieces of the puzzle). These bids should
be beneficial for both the buyer (total cost should decrease), and the bidder (bidder
should make a profit).
Our hypothesis is that providing this kind of “quantity support” the sealed-bid or semi-
sealed-bid auction would reach a more efficient outcome. The support tools should
also be considered fair by both the buyer and the sellers, because they try to maximize
the bidders’ profit, all the while decreasing the total cost to the buyer. Also, providing
support for the bidders – and thereby making bidding easier and less costly – the
auction would be more attractive, and more bidders would participate. More
competition should improve the buyer’s position, as she can expect to obtain the items
for a lower total cost.
The main methods used in this study are simulations and laboratory experiments. The
simulations were used to study the performance of the support tools. Through
simulations we could observe the efficiency of the final allocations, as well as the total
cost to the buyer. The laboratory experiment was used to study whether the simulated
results could be reproduced with human users. The laboratory experiments were also
used to observe bidders’ behavior in combinatorial auctions. Based on bidders’
behavior I identified different bidding strategies. In addition, I could draw conclusions
on how difficult a bidding environment is for inexperienced bidders, and how good the
usability of the user interface is.
12 A semi-sealed-bid auction is a sealed-bid auction in which the bidders know which of their own bids are among the provisional winners (= active) and which are not (= inactive).
90
IV DESIGNING AND TESTING THE QUANTITY SUPPORT
MECHANISM
7 THE QUANTITY SUPPORT MECHANISM13
The contribution of this chapter is to present the Quantity Support Mechanism
(QSM), a bidder support tool we developed for continuous, semi-sealed-bid
combinatorial auctions. I also provide an example auction to illustrate how the QSM
works in an auction. The auction mechanism we consider is such that bidders are free
to enter bids at any time in the auction, and the WDP is solved after every incoming
bid. In this auction, the set of provisional winners changes if the new bid decreased the
total cost to the buyer (in a reverse auction) by at least a predetermined decrement.
The decrement is public knowledge. The bidders receive information on their bid
status (whether they are provisional winners or not), but they do not know the contents
of the competitors’ bids. The auction has a predetermined closing time, but the time
will be pushed back if there is bidding activity in the last minutes of the auction. The
winning bidders receive the price indicated by their bid (or in forward auctions, pay
their bid price). We chose to consider an iterative, pay-your-bid auction mechanism,
because it resembles ideology of the English auction many bidders are familiar with
(see discussion in section 2.4).
The QSM has been designed for reverse auctions – and therefore the following
discussion will be from the reverse auction perspective – but it can easily be applied to
forward auctions. The purpose of the QSM is to suggest bids (price-quantity
combinations) to bidders who wish to become provisional winners. The QSM would
use the existing bids to solve for good complements, and then suggest these
complementing bids to the bidder without revealing the contents of the other bidders’
bids. Also, the idea is that the bid suggestions would be the best possible for the bidder,
i.e. bids which would maximize the bidder’s profit while having low enough a price so
that they would become winners at that time in the auction. Continuing the puzzle
13 Material in this chapter and section 8.1 has been published in Leskelä, Teich, Wallenius and Wallenius (2007).
91
analogy, the QSM solves for the size and shape of the missing piece in the puzzle. And
because by choosing different pieces (existing bids), different puzzles can be compiled,
the QSM chooses the one it thinks is the most profitable for the bidder.
7.1 The Quantity Support Problem
If the bidder could express her costs in a functional form, and if she would be willing to
disclose the cost function to the bid taker (or, if it could be arranged, to a neutral third
party), it would be fairly straightforward for the bid taker to solve for the bid that
maximizes the bidder’s profit. The quantity support problem for bidder m (QSPm)
would reduce to a standard mixed integer programming problem. The objective (5) of
the problem is to maximize the profit of bidder m by solving for the new price pm,new, the
vector Qm,new of item quantities qm,new,k and the values for the bid status variables xij. In
order for the new bid to become active (provisional winner), the current total cost to
the buyer C* is required to decrease by a predetermined decrement δ as a result of the
new bid (6), and the demand for each item dk must be fulfilled (7). The item quantities
qm,new,k should not exceed the bidder’s corresponding capacities amk (10)14. It is also
assumed that at most one bid per bidder can be active at a time (8), (9) to simplify the
bidding language. The formulation of the QSPm is thus:
14 We have included only these simple, per item capacity constrains in our formulation. Allowing the bidders to announce capacity constraints for combinations of items would increase the number of constraints in the formulation, reducing its readability. If desired, such more complex capacity constraints can be incorporated in the design.
92
(QSPm)
{ }
Kkqp
jix
Kkaq
Nix
njx
Kkdqqx
Cppxts
Qcp
knewmnewm
ij
mkknewm
n
j
ij
mmj
kknewm
N
i
n
j
ijkij
newm
N
i
n
j
ijij
newmmnewm
i
i
i
,...,10,
,1,0
,...,1
,...,11
,...,10
,...,1
*..
)(~max
,,,
,,
1
,,
1 1
,
1 1
,,
=∀≥
∀∈
=∀≤
=∀≤
=∀=
=∀≥+
−≤+
−
∑
∑∑
∑∑
=
= =
= =
δ
(5)
(6)
(7)
(8)
(9)
(10)
where ( )newmm Qc ,
~ is the cost function of bidder m, and as in the WDP (2), xij’s indicate
which bids are among the winners, and pij indicates the price and qijk the quantity of
item k in bidder i’s jth bid.
Note that even though we call our tool the “quantity support mechanism”, it also
suggests a price to attach to the quantities. The quantity support problem in (QSPm) is
presented for a price-only auction, but if we use the “pricing out” –method as in Teich
et al. (2001, 2006), also multiple attributes could be included in the bids. Also, if the
bids in the bid stream were disclosed to all bidders (an open-cry auction), each bidder
could formulate her own quantity support problem similar to (QSPm) replacing (.)~mc
with her own cost function (or an approximation of it).
However, it is not realistic that the QSPm as such could be applied into practice. It is
possible that the bidders are not able to express their costs in a functional form.
However, we do assume that the bidders are still capable of comparing the profitability
of different bids. Also, it is unlikely that they would be willing to disclose their cost
functions to the auction owner or even a neutral third party, even if they could specify
the functions. Therefore we need to find a way to approximate the bidders’ cost
functions, and preferably without having to ask for information from the bidders.
In this study we used a linear approximation of the bidders’ cost using the dual prices of
the demand constraints of the WDP as the variable cost parameters. The dual prices
93
can be interpreted as market prices for the items (see discussion on combinatorial
auction mechanisms in section 3.2.3.2 of the literature review), hence they can be
expected to reflect the underlying costs as well. Because in the integer programming
case there are no dual prices, we used the dual prices of the linear relaxation. This
means that the WDP (2) is solved again, but the binary constraints { }1,0∈ijx are
replaced with 10 ≤≤ ijx . The cost function ( )newmm Qc ,
~ in (5) is replaced with the linear
cost function
∑=
=K
k
knewmknewmm qQc1
,,, )(~ µ (11)
where µk is the dual price of the kth quantity constraint in the linear relaxation of the
WDP, and thereby the dual price for item k.
According to economic theory, firms have two kinds of costs: variable and fixed. We
have considered only the variable costs in our linear approximation. A fixed cost term
could easily be added, but as it is a constant, it would not affect the solution of the
maximization problem. The lack of the fixed cost element affects the value of the
objective function, but in this problem the approximated profit indicated by the
objective function is not interesting – only the allocation and bundle price are.
The resulting linear cost function has several limitations that need to be taken into
account. First of all, it cannot portray economies of scale. Thus, in case the bidders
experience economies of scale, the bidders’ costs for large bundles are systematically
overestimated. This would mean that there would be a bias in the QSP towards smaller
bids. However, since the objective is to maximize absolute profit (not relative), we
expect the very small bids to be eliminated anyway. Secondly, a linear cost function is
incapable of portraying economies of scope (= subadditive cost function). Because of
that the QSP has no incentive to try to bundle many items into the same bid. It is the
existence of economies of scope, which was the reason to organize a combinatorial
auction in the first place, hence we should assume the bidders’ true costs to exhibit
economies of scope. Therefore, one purpose of this study is to determine, whether a
linear approximation of the cost functions is good enough.
94
Thirdly, notice that the linear cost function estimate (11) is the same for all bidders,
hence it does not account for individual differences in the bidders’ cost functions.
Thereby the bid solutions of QSP are anonymous (i.e. the QSP gives the same
suggested bids regardless of the bidder). The only differences in the suggestions can
arise from the requirement that maximum one bid per bidder can be active at a time.
Thus, the bid suggestion for bidder m cannot be such that it would team up with
bidder m’s previous bids, whereas for any other bidder these bids can be teamed-up
with. We also recognize that there are differences between the bidders’ cost structures,
so the “anonymous” suggestion offered by the QSP may not be the best for all bidders.
Due to these limitations, it is possible that the solution of the QSP is not an acceptable
bid suggestion for any of the bidders. Therefore a shortlist of alternative solutions is
generated, and the shortlist alternatives are presented to the bidders together with the
solution from the QSP, and the bidder can choose the bid which is the most profitable
one for her. This can be done in two ways. First, we can go through the neighboring
pivots of the original quantity support problem. Pivoting in the integer case is
interpreted as solving the QSP over and over again, but each time forcing one status
variable xij with a zero value in the original QSP solution to assume the value “one”.
Hence, the number of pivots depends on the number of bids in the bid stream (as there
is one status variable for each bid), and the number of bids in the optimal combination
(those which assume the value “one” in the original QSP solution). The second
alternative is to solve the QSP over and over again, but this time setting different Qi’s to
zero to generate new combinations. If all the possible combinations are searched
through, there are 2K-2 combinations to go through, so in larger auctions it would not
be feasible. Oftentimes, though, the different pivots, as well as the different
combinations produce the same bid suggestion, so the number of non-identical items
on the shortlist hardly ever reaches the theoretical maximum. The QSP together with
the shortlist forms the core of the bidder support tool we call the QSM.
7.2 An Example of a Combinatorial Auction with the QSM
Consider a procurement situation, where a single buyer desires to buy a bundle of
items: 100 units of item A, 100 units of item B, and 200 units of item C. Her
95
reservation price for the entire package is $100,000. If there were other attributes, they
have been “priced out”. Let us further assume that there are three bidders: X, Y, and Z.
Finally, assume that the desired price decrement δ is equal to $3,000 from one bid to
the next.
Assume bidder X makes her first bid of 50 units of item A and 100 units of item B (but
no units of C) for a total bid price of $25,000. In vector notation, this bid is (50; 100; 0;
$25,000). Naturally, since this was the first and only bid so far, it is inactive (i.e. not
among the provisional winners), but it is entered into the bid stream. Next, assume that
bidder Y enters the following bid (100; 50; 200; $79,000). The bidder is informed that
her bid is not a provisional winner. Considered jointly with the bid of bidder X, they
would meet the quantity demand of the buyer. In fact they would exceed the demand
for items A and B, since bid X1 + bid Y1 = (150; 150; 200; $104,000). That would be
acceptable, except that the buyer’s reservation price ($100,000) is exceeded, making
the combination of bid X1 and Y1 not feasible. However, as previously, we retain bid Y1
in the bid stream. Next, bidder Z enters the following bid (100; 100; 0; $32,000). This
bid is also inactive. Bidder Z is informed. The other two earlier bids remain in the bid
stream with inactive status.
Assume that bidder X decreases the price on her original bid to $21,000. Now together
the bids of bidders X and Y become provisional winners. They are notified of their
changed status. Note that the new bid X2 contains the same quantities as X1 but for a
lower price. Thus we can drop X1 from the bid stream because each bidder can have
only one bid active simultaneously.
Bidder Z decides to use the quantity support tool to find an “active” bid. First, in order
to obtain the dual prices for the quantity constraints, we formulate and solve the LP
relaxation of the WDP:
96
10
10
10
200200
10010050100
10010010050..
000,32000,79000,21min
1
1
2
1
112
112
112
≤≤
≤≤
≤≤
≥
≥++
≥++
++
z
y
x
y
zyx
zyxts
zyx
(12)
where x2, y1, and z1 are the bid status variables. The solution of the problem is x2 = 0.5,
y1 = 1, z1 = 0. The dual prices are 0, 210, and 342.5 for the three quantity constraints.
Using these dual prices as the coefficients for the linear cost function we can formulate
the following quantity support problem. Denote the unknown price by pnew and the
unknown quantities by qnew,A, qnew,B, and qnew,C,. Here we assume that any bidder can have
only one active bid, so bid Z1 is deleted from the quantity support formulation. The
QSPz can then be formulated as
{ }1,0,
200200
10050100
10010050
000,97000,79000,21..
5.342210max
12
,1
,12
,12
12
,,
∈
≥+
≥++
≥++
≤++
−−
yx
qy
qyx
qyx
pyxts
qqp
Cnew
Bnew
Anew
new
CnewBnewnew
(13)
The suggested bid is (100; 100; 200; $97,000) with x2 = 0 and y1 = 0. The shortlist is
generated by forcing the bid status variables of inactive bids to assume the value “one”
in turn. Problem (13) is solved twice, first with the additional constraint x2 = 1, and
then with y1 =1. The shortlist consists of the following three bids: (100; 100; 200;
$97,000), (50; 0; 200; $76,000), (0; 50; 0; $18,000).
Bidder Z decides to accept the second bid from the shortlist, which is added to the bid
stream. The new bid Z2 becomes a provisional winner together with X2; Y1 becomes
inactive. The bidders X and Y are informed.
Next bidder Y requests a suggested bid. The linear relaxation of the WDP (similar to
(12)) is solved to obtain the dual prices. They are: 0, 210, and 380. The quantity
support problem is formulated as in (13), but this time Y1 is left out, and X2, Z1, and Z2
97
are included. The total cost to the buyer cannot exceed $94,000. The suggested bid is
(50; 0; 200; $73,000) with x2 = 1 and z1 = z2 = 0. The following shortlist is generated:
(50; 0; 200; $73,000), (0; 0; 200; $62,000), (50; 100; 0; $18,000).
Bidder Y decides to accept bid Y2 = (0; 0; 200; $62,000), which is then added to the bid
stream. Bid Y2 teams up with Z1, and Bidder X becomes inactive. The bidders are
informed about their new status. The auction continues along these lines until no
bidder is willing to place a new bid.
98
8 TESTING THE QUANTITY SUPPORT MECHANISM:
SIMULATION STUDIES
The contribution of this chapter is the analysis the results of the simulation studies
conducted to test the QSM, and the insights and deepened understanding of
combinatorial auctions obtained through the analysis. Especially the insights from the
second simulation study were important and led to new ideas and significant
improvements on the efficiency of the quantity support tools.
Simulations are a convenient way test solution algorithms and auction mechanisms.
Many researchers have used them in their research (e.g. Kelly and Stenberg, 2000,
Parkes, 2001, and Sandholm et al., 2005). The advantage of simulations over laboratory
experiments – which are another common research method – is that they are faster and
cheaper to set up, and it is much easer to test large-scale (say 30 bidders) auctions
through simulations. Therefore, we chose to test the QSM first with simulations. The
laboratory experiments with human subjects are presented in Chapter 12.
Two separate simulation studies were conducted to test the properties of the quantity
support mechanism. The auctions in both studies are reverse auctions. The
mechanism used is a first-price, semi-sealed auction, as used in the example auction in
the previous chapter.
The first study compares the QSM using dual prices in the objective function with a
quantity support mechanism, where the dual prices have been replaced with random
coefficients. The purpose of this study is to obtain validation for the use of dual prices.
The second simulation study studies the convergence properties of an auction where
the QSM is used. The main point of interest is the final auction outcome, i.e. what is
the total cost to the buyer, and how the items are allocated to the bidders.
8.1 First Simulation Study
The first simulation study consisted of three phases. In the first phase all the auction
parameter values were chosen and cost functions were created for the bidders. Also a
set of initial bids had to be generated into the bid stream because the QSM cannot be
used before there are some bids to team up with. In the second phase, all bidders used
99
the quantity support mechanism, and the most profitable bid from the shortlist for each
bidder was recorded. In this simulation study we were only looking at one step in the
auction, so no further bids were considered. In the third phase, the simulation results
were compared against two benchmark cases. In the first benchmark case, the QSM
was modified so that the dual prices in the objective function were replaced with
random numbers. This represented the “no information” case. In the second
benchmark case the approximate cost function is replaced by each bidder’s true cost
function in turn. This benchmark represented the ideal “perfect information” case, and
constituted the largest possible profit obtainable for any bidder at that point in the
auction. The profits obtained by bidders in phase two were then compared against the
two benchmark cases.
8.1.1 First Phase: Setting Up the Auctions
In any simulation study, some initial values have to be assumed. In this study we
needed to choose the number of bidders, number of items and the quantities
demanded of each item. The bidders needed to be assigned cost functions so that it
became possible to evaluate the profitability of the shortlist items offered by the QSM.
Also, a set of initial bids from the bidders had to be generated, because the QSM
cannot be used without some bids already in place in the bid stream.
8.1.1.1 The Cost Function
We wanted the cost function to be as simple as possible, but also we wanted it to
portray both economies of scale and scope. The use of combinatorial auctions is
justified in a situation in which there are synergies between the items. In reverse
auctions, synergies between items can be understood as synergies in the production
process of the items, i.e. economies of scope. Thus, it would not make sense to use a
cost function that would not allow for synergies in production.
The simplest form for a cost function exhibiting economies of scale is a linear function
with a fixed cost element:
C(qk) = Fk + ckqk (14)
100
where Fk is the fixed cost, and ck the per-unit cost of producing item k alone.
For the two-item case the cost function would take the form of
C(q1, q2) = F12 + c1q1 + c2q2 (15)
The existence of economies of scope in this framework implies simply that the
inequality of Equation (16) between the fixed cost parameters holds:
F12 < F1 + F2 (16)
The above presented multi-product cost function (Equation (15)) is very simplistic. It is
theoretically very restrictive, as it implies constant marginal costs and monotonically
decreasing average costs. The function is discontinuous at points when the level of one
or more outputs is zero, which makes it difficult to use in optimization problems. The
function is easy and convenient to use only in situations with relatively few products. It
can easily be seen that the number of different fixed cost elements increases rapidly as
the number of products increases. In the case of two products, there are only three
parameters F1, F2, and F12. With three products there are seven parameters: F1, F2, F3,
F12, F13, F23,, and F123, where Fij indicates the fixed cost of producing goods i and j.
When the number of products is increased to five there are already 31 fixed cost
parameters, with six products there are 63, and with seven products 127 parameters.
However, the simplicity of the function makes it intuitive and it is very flexible as it can
represent economies of scope of different magnitudes between different items – and
even diseconomies of scope between so items, if necessary. Thereby it is very appealing
in a theoretical framework such as ours.
Another reason for our choice of cost function was that there are not very good
alternatives available. Cobb-Douglas and CES (constant elasticity of scale) forms can
be used for multi-product cost functions, but they would have to be linearized before
they could be used in linear or integer programming problems. A commonly used form
for the cost function is the translog cost function (see Equation (17)), which is a
function of the output quantities (yi) and input prices (pj).
101
∑∑∑∑
∑∑∑∑
++
+++=
n
i
m
j
jiij
m
j
m
h
hjjh
n n
k
kiikj
m
ji
n
i
pypp
yypyC
lnlnlnln2
1
lnln2
1lnlnln
111
0
δγ
σβαα (17)
The translog function is more versatile, as it does not force homogeneity or constant
elasticity on the cost structure. Due to its flexibility, the translog cost function has been
popular in empirical studies which aim at estimating real world cost functions for some
firm or industry (see e.g. Murray and White, 1983, and Cho, 2003).
The translog cost function, however, is too complex for the purposes of this study. In a
simulation study it would be good to minimize the number of parameters to choose.
Thus, the inclusion of the input prices in the cost function is an unnecessary
complication. We chose to use the simple form of the cost function presented in
Equation (15). It is intuitive, easy to use, and sufficient for the purposes of this study,
where we only want to find out whether the QSM works under some circumstances.
8.1.1.2 Parameters
In order to reduce the sensitivity of the results to the initial values, we decided to vary
some of them. The number of bidders was fixed at 10, but the number of items to be
auctioned was either 3 or 5. The quantity demanded was 1000 for each item. The
variable cost parameters were drawn from the same uniform distributions in each
design. In the three-item auctions the variable cost parameters were drawn from the
range [30, 50] for c1, [40, 60] for c2, and within [60, 70] for c3. For the five-item auction
the variable costs fort items 1, 2 and 3 were drawn from the same range as in the three-
item auction, and for the additional items from the range [15, 45] for c4 and [20, 55] for
c5. The distributions for the variable cost parameters (and fixed cost parameters) were
the same for all bidders, so I have omitted the index indicating the bidder from the
notation.
The uniform distribution from which the fixed cost parameters (Fijk) were drawn had
two possible levels: “high” and “low”. The ranges for the fixed cost parameters were
chosen so that it was very likely that economies of scope would exist. This meant that
the lower bound of F12 was less than or equal to the sum of the lower bounds of F1 and
102
F2. A similar logic was applied to the upper bounds. It was also kept in mind that total
fixed cost should not decrease from the addition of a new product. Thus the lower
bound of F12 was set higher than the upper bounds of F1 and F2.
When the costs were low, the lower bounds ranged from 700 (for F3) to 2300 (F123) and
the upper bounds from 1000 (F3) to 3200 (F123) in the three-item auctions. Again, in the
five-item auctions the parameter ranges were the same for the first three items. The
lower bounds ranged from 700 (for F3) to 5500 (F12345) and the upper bounds from 1000
(F3) to 7500 (F12345). The exact ranges for all fixed cost parameters can be seen in
Appendix 1. When the costs were high, the lower bounds for the fixed costs ranged
from 5000 (F3) to 42000 (F12345) and upper bounds from 7000 (F3) to 50000 (F12345). The
new ranges can also be seen in Appendix 1. The higher fixed cost values were used to
test the effect of more pronounced economies of scope on the results. With the lower
level of fixed costs, the proportion of fixed costs in the total cost was only about 10%.
After the increase the proportion of fixed costs was almost 30%.
The initial bids in the bid stream were created so that each bidder was assumed to have
placed one bid. Thus, the number of bids in the initial bid stream equaled the number
of bidders. The quantities in the initial bids were drawn randomly from a uniform
distribution [100, 500], and rounded to the nearest 50. Some of the bid quantities,
however, were chosen to be zero in order to create some sparsity in the bid matrix. It is
realistic to assume that all bidders would not place bids for all products but a subset of
them. The level of sparsity was 20%. The constraint qnew,k ≤ 500 was added to the
quantity support problem to simulate the capacity constraints of the sellers. The
capacity constraint was set to simulate the fact that no bidder alone would be capable
of producing the whole demand. The bid price in the initial bid was the cost for the
bidder of producing that specific bundle, to which an initial mark-up of 30% was
added.
The simulation study consisted of four different experiment settings displayed in Table
4. Five replications of each setting were conducted.
103
Table 4 Design of the first simulation study
Experiment Items Bids Fixed Cost
I 3 10 Low (~10%) II 5 10 Low (~10%) III 3 10 High (~30%) IV 5 10 High (~30%)
8.1.2 Second Phase: Simulations
The simulations advanced as follows. First, the winner determination problem was
solved for the initial bid stream. The solution of the WDP determined the first set of
“active” bidders (or provisional winners), and the current lowest total cost for the buyer.
The linear relaxation of the same WDP was also solved to obtain the dual prices to be
used in the quantity support problem (QSP). Even though we knew the bidders’ cost
functions, we assumed that the auction owner solving the QSP would not know them.
The decrement δ with which the total cost to the buyer must decrease in the new
winning combination was set to 5% of the total cost in every simulation. The size of the
decrement generally has an impact on the convergence of auctions (as discussed in
section 5.1 on minor auction design issues). However, since we only study the first
incoming bid – and not the convergence – the choice of the decrement does not have
a major impact. Thus, different levels of the decrement were not tested.
With the decrement defined, the QSP could then be solved. The shortlist was
compiled through solving the QSP again and again adding the constraint xi = 1 for
each original non-basic variable (all xi for which xi = 0 in the WDP) in turn. The profit
for each bidder from each shortlist item was calculated as the difference between the
suggested bid price and the production costs for the suggested quantities. The largest
profit was recorded. In case all shortlist items produced a loss, the profit was set to zero
indicating the fact that the bidder would choose not to bid anything. The shortlist
items were evaluated also for the active bidders even though it would be more logical
to evaluate them only for inactive bidders. The reasoning behind this decision was that
active bidders may want to use the QSM to find out if they could obtain a bigger profit
104
than the 30% mark-up in their initial bids. The largest absolute profit for each bidder
was recorded15.
8.1.3 Third Phase: Benchmarks
Using the same initial data (cost function parameters and initial bids), two benchmark
cases for the QSM were solved. The first benchmark case repeated the procedure of
phase two with the exception that instead of dual prices, randomly chosen cost
parameters were used in the QSP. The parameters were chosen from a continuous
uniform distribution with the range [0, 250]. This interval was chosen because the dual
prices generally seemed to fall in the same range. The random parameters did not
utilize any information available on the cost functions of the bidders, and therefore
represented the extreme case of “no information”.
In the second benchmark – the case of “perfect information” – the whole
approximated cost function in the QSP was replaced by each bidder’s true cost
function in turn. The IP formulation of the profit maximizing problem is not trivial
due to the discontinuous cost function. The formulation is presented in Appendix 2. In
the perfect information case there is no shortlist to be created because the optimum is
found directly for each bidder.
8.1.4 Results of the First Simulation Study
We were primarily interested in comparing the profits obtained by the bidders using
quantity support to profits in the two benchmark cases. On the one hand we wanted to
know, whether the QSM using dual prices performed better than the QSM with
random cost parameters, and on the other we wanted to find out, how close to the best
possible profit the QSM could get. Thus, when presenting the results I will mainly
focus on these comparisons. However, as a sideline I have also looked at the bidders’
mark-ups from the most profitable bids to see how they change from the 30% set in the
initial bids. The mark-ups are considered only for the bids generated by the dual price
QSM.
15 Using the largest profit as bidders’ decision rule is only one alternative. Other alternatives would be the largest mark-up (ratio of profit to total cost), or largest turnover (price) which is still profitable.
105
8.1.4.1 Profits
The most interesting aspect concerning the bidders’ profits from the use of the QSM is
to look at how close to the maximum profit (or perfect information case) we could get.
Another point of interest is to compare the performance of random parameters (the no
information case, or “random support”) to the quantity support. The performance
measure used to compare the outcomes was the percentage of maximum profit
available that each bidder could get with both quantity support and random support:
*
i
iD
Deperformancπ
π= (18)
where iDπ is the profit of the most profitable bid suggested to bidder i by the QSM
using dual prices, and *
iπ the optimal profit of the perfect information case calculated
using bidder i’s true cost function. The performance of random support, performanceR,
is calculated by substituting into the numerator of (18) the profit of the most profitable
bid suggested to bidder i by the QSM using random cost parameters.
The performance indicators for QSM with dual prices (D) and “random support” (R)
for the first three experiments (see Table 4 for the details of experiment designs) are
presented in Table 5 (Experiment I), Table 6 (Experiment II) and Table 7
(Experiment III).
Table 5 Results of Experiment I: Performance indicators of the dual price quantity support (D) and random support (R)
Bidder 1 Bidder 2 Bidder 3 Bidder 4 Bidder 5 Bidder 6 Bidder 7 Bidder 8 Bidder 9 Bidder 10
D 0,86 0,87 0,76 0,14 1,00 0,95 0,87 0,95 0,76 0,00
R 0,59 0,57 0,39 0,33 0,72 0,62 0,66 0,36 0,61 0,33
D 0,88 1,00 0,87 0,88 0,92 1,00 0,86 0,00 0,90 0,91
R 0,27 0,55 0,68 0,59 0,60 0,58 0,66 0,67 0,07 0,55
D 0,00 0,60 0,73 0,83 0,74 0,79 0,00 0,84 0,82 0,75
R 0,82 0,72 0,74 0,73 0,75 0,74 0,93 0,77 0,73 0,75
D 0,92 0,88 0,00 0,79 0,91 0,86 0,42 0,88 0,95 0,91
R 0,42 0,41 0,12 0,40 0,44 0,38 0,41 0,43 0,41 0,26
D 0,76 0,83 0,92 0,00 0,87 0,83 0,66 0,87 0,86 0,00
R 0,65 0,76 0,82 0,85 0,75 0,16 0,72 0,80 0,83 0,00
4 4 3 3 4 5 3 4 5 4
Replication
# cases
where D≥R
1
2
3
4
5
106
Table 6 Results of Experiment II: Performance indicators of the dual price quantity support (D) and random support (R)
Bidder 1 Bidder 2 Bidder 3 Bidder 4 Bidder 5 Bidder 6 Bidder 7 Bidder 8 Bidder 9 Bidder 10
D 0,78 0,73 0,78 0,49 0,56 0,72 1,00 0,79 0,73 0,75
R 0,38 0,26 0,15 0,00 0,33 0,21 0,22 0,24 0,40 0,29
D 0,90 0,74 0,80 0,78 0,74 0,00 0,77 0,80 0,82 0,80
R 0,73 0,62 0,57 0,00 0,67 0,00 0,69 0,60 0,00 0,58
D 0,85 0,65 0,80 0,00 0,85 0,78 0,76 0,00 0,84 0,86
R 0,24 0,20 0,26 0,00 0,00 0,23 0,30 0,00 0,38 0,35
D 1,00 0,86 1,00 1,00 1,00 0,00 0,94 1,00 1,00 0,00
R 0,34 0,28 0,29 0,05 0,30 0,00 0,31 0,39 0,39 0,44
D 0,98 0,76 0,84 0,76 0,00 1,00 0,91 1,00 0,47 0,00
R 0,34 0,34 0,42 0,00 0,37 0,17 0,36 0,37 0,33 0,11
5 5 5 5 4 5 5 5 5 3
Replication
# cases
where D≥R
1
2
3
4
5
Table 7 Results of Experiment III: Performance indicators of the dual price quantity support (D) and random support (R)
Bidder 1 Bidder 2 Bidder 3 Bidder 4 Bidder 5 Bidder 6 Bidder 7 Bidder 8 Bidder 9 Bidder 10
D 0.50 0.55 0.47 0.45 0.48 0.00 0.44 0.46 0.48 0.51
R 0.26 0.23 0.23 0.24 0.25 0.00 0.31 0.30 0.32 0.51
D 0.98 0.53 0.68 0.55 0.00 0.73 0.62 0.65 0.66 0.73
R 0.98 0.50 0.68 0.55 0.31 0.73 0.62 0.65 0.66 0.73
D 0.52 1.00 0.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00
R 0.06 0.22 0.28 0.17 0.24 0.25 0.19 0.28 0.18 0.28
D 0.63 0.57 0.59 0.54 0.59 0.62 0.57 0.31 0.64 0.53
R 0.67 0.55 0.66 0.61 0.40 0.58 0.49 0.65 0.57 0.64
D 0.83 0.84 0.73 0.89 0.85 0.86 0.68 1.00 0.87 0.83
R 0.68 0.76 0.58 0.73 0.61 0.00 0.68 0.79 0.81 0.71
4 5 3 4 4 5 5 4 5 4
# cases
where D≥R
2
3
4
5
Replication
1
The results were very promising in general. The QSM using dual prices categorically
found profits that were above 70 % of the maximum, as can be seen from the tables.
Also, in some cases the bidder could actually obtain the maximum profit with the help
of the dual QSM (indicated by “1” in the tables). This means that the optimal bid for
the bidder in question was on the shortlist. The profits obtained with the random
approach were often lower, and varied more. In pairwise comparisons, the dual price
approach (D) performed as well as or better than the random cost approach (R) 86% of
the time. I also tested the statistical significance of the results using a pairwise t-test.
The p-values (two-tailed) for the three respective tables were highly significant (0.007,
0.0000, 0.0001).
In Experiment IV (5 products, 10 bids and higher fixed costs), the quantity support tool
using dual prices did not perform as well as the one with random cost parameters. The
performance indicators for dual price quantity support and random support in
Experiment IV can be seen in Table 8.
107
Table 8 Results of Experiment IV: Performance indicators of the dual price quantity support (D) and random support (R)
Bidder 1 Bidder 2 Bidder 3 Bidder 4 Bidder 5 Bidder 6 Bidder 7 Bidder 8 Bidder 9 Bidder 10
D 0,70 0,50 0,67 0,69 0,69 0,69 0,60 0,70 0,74 0,67
R 0,97 1,00 1,00 1,00 1,00 1,00 1,00 1,00 0,39 0,82
D 0,47 0,00 0,72 0,77 0,76 0,79 0,69 0,87 0,70 0,69
R 0,90 0,74 0,81 0,88 0,88 0,84 0,65 0,00 0,81 0,74
D 0,00 1,00 0,85 0,00 1,00 0,91 * 1,00 0,92 0,72
R 0,46 0,57 0,33 0,52 0,65 0,58 * 0,55 0,61 0,61
D 0,87 0,86 0,70 0,00 0,87 0,84 0,90 0,00 0,87 0,87
R 0,81 0,96 0,89 1,00 0,95 0,88 0,98 0,00 0,99 0,96
D 0,33 0,14 0,00 0,34 0,34 0,47 0,08 0,34 0,00 0,32
R 0,71 0,87 0,67 0,68 0,29 0,82 0,66 0,68 0,70 0,72
1 1 1 0 2 1 1 3 2 1
* Lindo could not find a solution
Replication
# cases
where D≥R
3
4
5
1
2
The better performance of the random support is also statistically significant (p-value
0.003). However, as can be seen from Table 8, the dual price approach is still
oftentimes reasonably good generating profits well above 60% of the maximum. The
random parameters simply worked even better.
One must keep in mind here, though, that the random parameters were chosen from
the interval within which the dual prices varied. In reality, only one or the other
approach would be used. So, if we were using random parameters alone, we would not
know the range within which the dual prices varied. We reproduced Experiment IV
with random cost parameters, but this time chosen from the range [0, 2000]. The
random parameters worked poorly producing very small profits categorically (see Table
9).
Table 9 Results of Experiment IVb: Performance indicators of the random support (R) with parameters drawn from the range [0, 2000]
Bidder 1 Bidder 2 Bidder 3 Bidder 4 Bidder 5 Bidder 6 Bidder 7 Bidder 8 Bidder 9 Bidder 10
D 0.70 0.50 0.67 0.69 0.69 0.69 0.60 0.70 0.74 0.67
R 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
D 0.47 0.00 0.72 0.77 0.76 0.79 0.69 0.87 0.70 0.69
R 0.03 0.00 0.12 0.11 0.12 0.05 0.11 0.06 0.11 0.11
D 0.00 1.00 0.85 0.00 1.00 0.91 * 1.00 0.92 0.72
R 0.00 0.02 0.21 0.00 0.25 0.04 * 0.25 0.22 0.23
D 0.87 0.86 0.70 0.00 0.87 0.84 0.90 0.00 0.87 0.87
R 0.42 0.40 0.25 0.00 0.42 0.28 0.45 0.00 0.00 0.43
D 0.33 0.14 0.00 0.34 0.34 0.47 0.08 0.34 0.00 0.32
R 0.36 0.00 0.36 0.01 0.02 0.44 0.35 0.35 0.01 0.35
4 5 4 5 5 5 4 4 4 4
* Lindo could not find a solution
5
# cases
where D≥R
Replication
1
2
3
4
The performance of the random parameters approach understandably seems to depend
on the chosen range of the cost parameters. The approach of generating random
parameters is beautiful in its simplicity, but their use is complicated by the fact that an
108
unsuitable range can significantly deteriorate the results. Thereby, we feel that using
dual prices is a better and more robust approach. It is also fairly simple, even though an
additional (linear) optimization problem has to be solved each time someone wishes to
use the quantity support tool.
8.1.4.2 Mark-Ups
An interesting detail related to the mark-ups (profit divided by production cost) in the
bidders’ bids can be observed in the results. Remember that the mark-up percentage in
the initial bids was set at 30%. One would assume that the mark-ups would have to
decrease in order for the new bids to become active. Interestingly, the profits generated
by the dual price QSM resulted in higher than 30% mark-ups in over 80% of the cases.
The mark-ups in the four experiments are presented in Tables 10-13. The mark-ups
correspond to the dual price QSM bids for which the profit ratios are presented in
Tables 5-8. In some cases the profits were actually larger than costs, i.e. mark-up was
over 100%. This demonstrates the benefits that can be obtained through searching for
combinations that team up well together. When the auction progresses, the mark-ups
will naturally have to start to decline.
Table 10 Mark-ups in Experiment I
Bidder 1 Bidder 2 Bidder 3 Bidder 4 Bidder 5 Bidder 6 Bidder 7 Bidder 8 Bidder 9 Bidder 10
Replication 1 0.29 0.31 0.41 0.08 0.35 0.36 0.31 0.33 0.32 0.00
Replication 2 0.37 0.54 0.51 0.46 0.44 0.43 0.44 0.00 0.42 0.37
Replication 3 0.00 0.42 0.50 0.56 0.51 0.34 0.00 0.49 0.60 0.46
Replication 4 0.52 0.71 0.00 0.62 0.64 0.80 0.59 0.62 0.50 0.65
Replication 5 0.44 0.47 0.48 0.00 0.44 0.55 0.42 0.53 0.58 0.00
Table 11 Mark-ups in Experiment II
Bidder 1 Bidder 2 Bidder 3 Bidder 4 Bidder 5 Bidder 6 Bidder 7 Bidder 8 Bidder 9 Bidder 10
Replication 1 0.36 0.36 0.31 0.46 0.55 0.45 0.39 0.42 0.29 0.31
Replication 2 0.42 0.68 0.38 0.43 1.11 0.00 0.50 0.64 0.35 0.49
Replication 3 0.51 0.53 0.77 0.00 0.99 0.65 0.62 0.00 0.35 0.93
Replication 4 1.05 0.89 0.98 1.02 1.14 0.00 0.86 1.10 1.11 0.00
Replication 5 0.82 0.59 0.58 0.59 0.00 0.85 0.63 0.70 0.24 0.00
Table 12 Mark-ups in Experiment III
Bidder 1 Bidder 2 Bidder 3 Bidder 4 Bidder 5 Bidder 6 Bidder 7 Bidder 8 Bidder 9 Bidder 10
Replication 1 0.80 0.88 0.47 0.59 0.75 0.00 0.66 0.70 0.86 0.33
Replication 2 0.44 0.33 0.54 0.40 0.00 0.43 0.45 0.46 0.45 0.46
Replication 3 0.31 0.53 0.00 0.59 0.45 0.65 0.61 0.51 0.45 0.58
Replication 4 0.38 0.18 0.31 0.27 0.22 0.27 0.24 0.38 0.29 0.18
Replication 5 0.48 0.47 0.46 0.42 0.51 0.55 0.45 0.48 0.54 0.55
109
Table 13 Mark-ups in Experiment IV
Bidder 1 Bidder 2 Bidder 3 Bidder 4 Bidder 5 Bidder 6 Bidder 7 Bidder 8 Bidder 9 Bidder 10
Replication 1 0.59 0.36 0.68 0.39 0.67 0.55 0.64 0.73 0.41 0.60
Replication 2 1.06 0.00 1.09 0.96 1.08 1.00 1.04 0.95 1.03 0.97
Replication 3 0.00 0.67 0.69 0.00 0.63 0.64 0.63 0.67 0.56 0.47
Replication 4 0.93 0.92 1.65 0.00 1.08 0.82 1.21 0.00 1.07 0.82
Replication 5 0.41 0.26 0.00 0.44 0.35 0.44 0.26 0.42 0.00 0.35
8.2 Second Simulation Study16
After concluding in the first simulation study that the dual price quantity support was
helpful, and better than random support, we wanted to test the QSM further. In the
first study we only considered the first bid added to the initial bid stream with the help
of the QSM. While that is an indicator of the QSM’s ability to find good
complementing bids, it still does not tell us anything about the final outcome of the
auction. Thus, in the second simulation study we wanted to look at the auction
outcomes. Our primary interest was in finding out what the buyer’s final cost ended up
being, and how close to an efficient allocation we could get with the help of the QSM.
In addition, we wanted to examine more thoroughly the sensitivity of the auction
outcome to the chosen input parameters. Our secondary interest was in studying the
speed of convergence, i.e. how long it took to reach the final allocation. The
convergence speed tells us the time the participants have to spend in the auction, and
is therefore an indication of the time cost involved. A mechanism that ultimately
results in an efficient or nearly efficient outcome may not be popular, if it takes too
long to reach to outcome. Therefore, the length of the auctions should not be ignored
in the testing phase.
The auction setting and mechanism in the second auction study was essentially the
same as in the first one. The second simulation study resembled the first one also in the
sense that it also consisted of three phases: defining the input parameters, the actual
simulation phase, and defining (and calculating) the benchmark cases. The biggest
difference between the two studies was naturally the fact that in the second study the
auctions are run all the way to the closing, but there were also some differences in the
16 Material presented in this section is based on joint work with Valtteri Ervasti (Ervasti and Leskelä, 2009, forthcoming in European Journal of Operational Research).
110
values of the simulation parameters, choice of parameters to vary, and in the way the
shortlist was compiled.
8.2.1 Phase One: Parameters
The logic by which the input parameters were created in the second study resembled
the first simulation study in many respects. The demand was the same across items;
dk = 600 for all k = 1, …, 5. The number of bidders was either 15 or 30. The bidders’
maximum capacity was again 50% of total demand, i.e. 300. However, this time we also
tested the effect of unequal capacities. In the unequal capacities case bidders had a
maximum capacity aik of 300 (with 50% probability), 225 (with 25% probability) or 150
(with 25% probability). The bidders’ capacities were defined separately for each item.
Thus, the bidders’ capacities varied from item to item. In the second simulation we
tried to improve the robustness of our results by increasing the number of replications
to 50. The buyer required that the total cost reduces by 2% each iteration. Thus, δ in
the QSP is 0.02C*. The size of the decrement affects the auction in two main ways.
The smaller the δ, the closer to the efficient production cost the total cost in the
auction can potentially go. However, the smaller the decrement, the slower is the
convergence of the auction. We chose 2% so that the auction would take long enough
that there would be relatively many bids in the auction for the QSP to use, and still not
have the auction take too long. Also, changing the decrement to 1% did not change the
results significantly (basically, the auction only continued one step further), hence we
decided not to vary the decrement in our final design.
Thus, the first two simulation design parameter values varied in the experiment were
number of bidders (15 or 30), and the bidders’ capacities (equal and unequal).
8.2.1.1 Cost Function
We used the same form for the cost function as in the first study (see Eq. (15)). Thus,
the cost function limited the number of items that could be handled easily. We chose
to use 5 items in every simulation. We made the choice based on pilot studies, which
indicated that changing the number of items from 3 to 4 or 5 did not change the results
significantly.
111
The cost function parameters were chosen again from uniform distributions, and the
distributions were the same for all bidders (i.e. symmetric situation). Thus, in this
section I have omitted the index i indicating the bidder from the notation. The variable
cost parameters ck were chosen from the uniform distribution [53.33, 66.67], with
E(ck) = 60 for all five items. The fixed cost parameters had again two different levels,
but this time the difference was in the degree of economies of scope, and not in the
proportion of the fixed costs of total cost. We denote these two cases as “normal”
economies of scope, and “large” economies of scope. These labels should not be taken
too literally, though. For simplicity, all items were treated identically, i.e. the
distribution for the fixed costs was dependent only on the number of items in the
combination. Normal economies of scope for a combination L was defined as
)(5.0)()( \ kL FEFEFE kL += (19)
so the inclusion of an additional item increases the expected fixed cost by 50% (and not
100% as in a linear case). The “large” economies of scope were defined in a similar
fashion as “normal” economies of scope, but the multiplier used was 0.4 instead of the
0.5 used in (19). The expected value for the single-item fixed costs was set so that the
expected proportion of fixed costs over total cost in a single-item bid at full capacity (=
300) would be 50%. The expected value of variable costs is .000,1830060 =× Thus,
E(Fk) = 18,000. The spread of the distribution around the mean was designed so that
the upper limit would be 25% higher than the lower limit. This translates to a
maximum of 25% cost advantage of one bidder over another. The ranges for the fixed
cost parameters with “normal” and “large” economies of scope are presented in Table
14 and Table 15.
Table 14 Ranges for fixed cost parameters in “normal” economies of scope
L Lower Limit Mean Upper Limit
1 item 16,000 18,000 20,000
2 items 24,000 27,000 30,000
3 items 32,000 36,000 40,000
4 items 40,000 45,000 50,000
5 items 48,000 54,000 60,000
112
Table 15 Ranges for fixed cost parameters in “large” economies of scope
L Lower Limit Mean Upper Limit
1 item 16,000 18,000 20,000
2 items 22,400 25,200 28,000
3 items 28,800 32,400 36,000
4 items 35,200 39,600 44,000
5 items 41,600 46,800 52,000
Thus, the third simulation design parameter that was varied in the study is the size of
the economies of scope (normal or large).
8.2.1.2 Initial Bids
In the second simulation study we wanted to generate the initial bids “intelligently”. In
the first study we simply drew the bid quantities randomly form a uniform distribution,
and then randomly forced some quantities to zero. Intuitively, it would make more
sense to have the bid quantities depend on the cost functions. Bidders in any auction
should have some idea on their cost level or the general cost level of the industry even
if they do not have exact information. We designed three different ways to generate the
bid quantities: one based on the evaluation of variable costs (“Bid1”), one based on
fixed costs (“Bid2”), and one based on the combination of fixed and variable costs
(“Bid3”). In each case, the price for each bid was obtained by calculating the
production cost for the bundle, and adding a 20% mark-up on top of the cost.
In Bid1, bidders’ variable costs were compared to the expected value of the distribution
of the variable costs. If cik < E(cik), the bid quantity qi1k (quantity of item k in bidder i’s
first bid) was set to the bidder’s maximum capacity of item k. Otherwise qi1k = 0. For
example, assume that the variable costs of bidder B1 for the five items were (55.9, 56.9,
61.0, 63.2, 57.3) respectively. The expected value of the distribution, E(cik) = 60. Thus,
bidder B1’s variable costs are less than average for items 1, 2, and 5. If B1’s capacities
were (300, 150, 150, 225, 300), the bidder’s initial bid quantities according to Bid1
would be (300, 150, 0, 0, 300).
In Bid2, a similar comparison was done with the fixed cost parameters. All the 31 fixed
cost parameters were compared to their expected values, and the one that was
proportionally most below the expected value was chosen. The bid quantities of the
items corresponding to the most advantageous fixed cost combination were set to
113
maximum capacity. Other qi1k = 0. For example, bidder B1’s fixed cost for items 2, 4
and 5 were 33,024 under normal economies of scope, the ratio to the expected fixed
cost (= 36,000 for three items) would be 0.917. If this were the smallest ratio among all
31 comparisons, the bidder’s initial bid quantities according to bid2 (assuming same
capacities as in the previous example) would be (0, 150, 0, 225, 300).
In Bid3, the comparisons considered the total production costs (fixed cost + variable
cost). Using the cost function parameters cik and FiL, we calculated the production cost
for each item combination (in total 31 combinations including the single items) at
maximum capacity of bidder i. The reference point was the expected total cost of
producing each of the 31 combinations at maximum capacity of bidder i. More
formally, the reference point is:
∑∈
+=Lk
ikikLLi acEFECE )()()( , (20)
where E(FL) is the expected fixed cost of combination L (see Table 14 and Table 15),
aik is the maximum capacity of bidder i for item k, and E(cik) = 60 for all i and all k. The
bidders’ actual fixed costs and variable costs are substituted in (20) to obtain the actual
cost, which is then compared to (20). The combination for which the ratio of actual
total cost to the expected cost, Ci,L / E(Ci,L), was the lowest, was chosen. Again, the
quantities in the most advantageous combination were set to maximum capacity, and
other quantities were set to zero. Continuing the example from above, assume that B1’s
actual fixed cost F1,245 = 34,687 and variable costs the same as earlier. Then B1’s actual
total cost for items 2, 4 and 5 would be 34,687 + 56.9×150 + 63.2×225 + 57.3×300 =
74,632. expected cost for the bidder B1’s bid for items 2, 4 and 5 at full capacity would
be 36,000 + 60 ×(150+225+300) = 76,500. The ratio of these two costs is 0.976, and if
it were the smallest ratio among all 31 ratios, bidder B1’s initial bid quantities would
again be (0, 150, 0, 225, 300).
Thus, fourth simulation design parameter to be varied in the study was the way in
which the initial bid stream was created (Bid1, Bid2, Bid3).
114
8.2.2 Phase Two: Simulations
For the simulation phase, the auction process is divided into four steps and presented
as an algorithm:
Step1: Winner Determination. The WDP (Eq.(2)) is solved for the set of bids in the
bid stream (in the beginning the set of bids in the bid stream are the initial bids). Based
on the solution, a set of bidders become provisional winners, and the rest are inactive.
Proceed to Step2.
Step2: Quantity Support Mechanism. The WDP of Step1 is solved again in the
linearized form creating a vector of dual prices. The QSP presented in section 7.1 is
solved using these dual prices as variable costs in the objective function.
Next, the shortlist is created. In this study, we used two different ways to create the
shortlist. The one called “full” shortlist contains the original solution of the QSP, and
the solutions of the QSP with one of the following additional constraints in turn:
xij = 1 in turn for all xij = 0 in QSP0
Lkqk ∈∀= 0 , for each subcombination L of the set of all items M (21)
The “express” shortlist only contains constraints of the latter type. The shortlist type was
the fifth (and last) simulation design parameter to be varied in the study. Proceed to
Step3.
Step3: Selection of Bidder and Bid. From the set of inactive bidders determined in
Step1 a bidder i is chosen randomly as the bidder “using” the quantity support. The
cost function of bidder i is used to evaluate the bids on the shortlist. The bid for which
the profit is the largest – i.e. )(max ,,,
∑∈
+−Lk
rkikLirqp
qcFprr
, where L is any subset of items,
pr is the price of the rth, shortlist item, and qk,r is the quantity of item k in the rth shortlist
item – is selected. If profit is nonnegative17, the bidder places the bid, and it is added
into the bid stream. Move back to Step1. If all profits are negative, another bidder from
17 We used zero profit as the limit for an acceptable bid. The lowest acceptable profit is only a matter of normalizing, and it is customary in economics to assume that “normal” profit is already included in the cost function (cf. “economic cost”).
115
the set of inactive bidders is chosen, and she evaluates the short list items. If no bidder
has accepted the suggested bids on the shortlist, and there are no inactive bidders left,
move to Step4.
Step4: Auction Ends. No inactive bidder accepted a bid suggestion made by the
QSM. Note that we assumed that after the initial bids are entered, no new bids are
entered outside the suggestions of shortlist. Thus, no new bids enter the auction, and
the auction ends. The provisional winners from Step1 become the actual winners, and
the total cost to the buyer equals the value of the objective function in the solution of
the WDP.
8.2.3 Benchmark Cases
In the second simulation study, we used three benchmark cases. As in the first study,
we had one benchmark case to represent the situation with less information (and
support), and one to represent the case of perfect information (the “first best” solution).
Also, as in the first simulation study, the “less information” case is studied as an
alternative to the QSM, and the “perfect information” case is what the two other cases
are compared against. The third benchmark case we used was an auction in which the
items were auctioned off individually, i.e. bids on combinations of items could not be
placed (the “non-combinatorial” case).
The case of less information means that the QSM is not available for the bidders.
Instead, in Step 2, they can only use the “suggested price” tool proposed by Teich et al.
(2001 and 2006). Teich et al. have presented the suggested price tool in the context of
multi-attribute auctions, but the idea is directly transferable to the multiple-item
setting. The suggested price tool in a combinatorial auction returns the price that will
make a given item combination a provisional winner. The bidder needs to specify the
quantities bk of each item beforehand. The suggested price problem for bidder m
(SPPm) can also be formulated as an integer programming problem:
116
(SPPm)
{ }1,0
,...,11
,...,10
,...,1
..
max
1
1 1
*
1 1
∈
=∀≤
=∀=
=∀≥+
−≤+
∑
∑∑
∑∑
=
= =
= =
ij
n
j
ij
mmj
kk
N
i
n
j
ijkij
new
N
i
n
j
ijij
new
x
Nix
njx
Kkdbqx
Cppxts
p
i
i
i
δ
(22)
(23)
(24)
(25)
(26)
We are assuming that the auction mechanism allows at most one bid from each bidder
to be active, and thus the bid status variables xmj for the bidder’s earlier bids are forced
to zero. The suggested price tool does not convey any information the bidder could not
obtain by herself with trial and error. However, it would be very impractical to have to
place a large number of bids with decreasing prices to discover the price that would
make the bid active. Thus, the suggested price tool is a welcomed convenience that
expedites the auction.
What are then the quantities bk that the bidder enters into the suggested price tool in
the simulation? In our simulations we assume that the bidder anchors on her initial
bid. She first asks for the suggested price for the initial bid, and then adjusts the bid
quantities and asks for new suggested prices. The additional combinations she
considers are all the subsets of the original bid (on item level), and a bid that is 50% of
the original bid. For example, assume that the bidder’s initial bid had the quantities [0,
300, 150, 0, 225]. She would then ask for a suggested price for the following set of bids:
[0 300 150 0 225]
[0 0 150 0 225]
[0 300 0 0 225]
[0 300 150 0 0]
[0 0 0 0 225]
[0 0 150 0 0]
[0 300 0 0 0]
[0 150 75 0 112.5]
117
In Step3 the bidder evaluates the resulting price-quantity combinations against her cost
function, and places a new bid, provided that at least one combination produces a
nonnegative profit.
The “perfect information” case in the second study was the efficient allocation of the
items between the bidders. By efficient I mean the allocation that minimizes the total
production cost of the bundle demanded by the buyer. When solving for the efficient
allocation, the bidders’ cost functions were assumed to be known, so the true optimum
could be discovered. The conceptually simple optimization problem is in fact quite
difficult (and lengthy) to formulate due to the discontinuous cost function, so the
formulation is presented in Appendix 3. The formulation is based on the same logic as
the optimum quantity support bid calculation in the first simulation study (Appendix
2).
In the non-combinatorial case we solved for the efficient allocation in a situation in
which each item was auctioned in a separate, multiple-unit auction. To simplify
matters we assumed that the winners of each auction would be the bidders with the
lowest production cost for that particular item that can fulfill the demand. Thus, no
strategic bidding that would consider the possible outcomes of the auctions of the other
items was taken into account. We also assumed that the outcome of the single-item
auctions would be efficient, and that there would be enough competition to drive the
prices close to the production costs (or normal profit). The comparison of the auctions
was then done based on the final cost to the buyer and not based on the production
costs as was done with the two other benchmark cases.
8.2.4 Results of the Second Simulation Study
In the simulation design there were five design parameters that we varied in the study:
number of bidders (15 or 30), bidders’ capacities (equal or unequal), the magnitude of
economies of scope (normal or large), the way the initial bid stream was compiled
(Bid1, Bid2 or Bid3), and the way the shortlist was compiled (full or express). Thus,
there were 48 different auction designs, and with 50 replications of each design, the
total number of simulated auctions was 2400. Each of the 2400 auctions was run
though first with price support and then with quantity support, and finally the efficient
118
allocation for each auction was solved. Note that the price support auctions do not use
the shortlist, which is part of the QSM. Hence, there were actually only 24 different
designs for the price support auctions, but still all the 2400 auctions were run through
with price support as well. Conducting price support auctions were for all “full” and
“express” shortlist auctions allowed for pairwise comparison with the corresponding
quantity support auctions.
In this second simulation study, our primary interest was in examining the final
outcome of the auctions. Thus, we studied the outcomes of the auctions from many
perspectives. We studied both the total cost to the buyer as well as the efficiency of the
final allocation. We contrasted the final outcomes of the quantity support auctions and
the price support auctions with the efficient allocation. The final outcomes of the
quantity support auctions were also compared against the outcomes of the individual,
non-combinatorial auctions. In addition, we studied the effects of the initial parameters
on the auction outcomes. Besides total cost to the buyer and allocative efficiency, we
were interested in the speed of convergence in the auctions and the usefulness of the
shortlist.
8.2.4.1 Total Cost to the Buyer
The total cost to the buyer at end of the auction is the primary concern of the buyer
(auction owner); it is only us researchers and perhaps the government as a buyer who
are interested in the efficiency of the final allocation. The buyer wants to minimize the
price she has to pay, and she does not care whether the allocation is efficient or not.
Naturally, the more efficient the allocation, the lower the total cost can potentially go.
However, even if the final allocation were efficient, the cost to the buyer can be high if
there is not much competition and the mark-ups remain high even at the end of the
auction. Hence the buyer will want the auction to be relatively efficient, but only as a
means to an end. If the QSM does not produce outcomes with prices acceptable to the
buyer, she will not want to use the mechanism. Thus, in our simulation study we were
interested in finding out how the use of the QSM affected the total price paid by the
buyer.
119
Table 16 contains the ratios of the average total cost to the buyer from the winning
allocations and the efficient production cost (theoretically the lowest possible cost to
the buyer) in quantity and price support auctions. For example, a ratio of 1.070 means
that the total cost to the buyer is 7% above the efficient production cost. The cost to the
buyer is averaged over all the simulation settings in which the parameter considered
has the same value. For example, there were 1200 auctions with 15 bidders, and 1200
auction with 30 bidders, and 800 auctions with Bid1, Bid2 and Bid3 respectively.
Table 16 Total cost to the buyer as ratio of efficient production cost
Quantity Support Price Support Mean St.dev Mean St.dev
15 1.070 0.054 1.239 0.120 # of bidders
30 1.054 0.039 1.157 0.085 Normal 1.062 0.048 1.196 0.111
Econ. Of scope Large 1.062 0.048 1.200 0.113 Bid1 1.065 0.051 1.168 0.098 Bid2 1.061 0.048 1.164 0.091 Initial bids
Bid3 1.060 0.045 1.261 0.117 Equal 1.025 0.014 1.128 0.069
Capacities Unequal 1.099 0.042 1.267 0.103 Full 1.058 0.042 1.196 0.114
Shortlist Express 1.066 0.053 1.199 0.110
ALL 1.062 0.048 1.198 0.112
As can be read from Table 16, the total cost to the buyer is much lower when quantity
support is used. The price the buyer has to pay is on average only 6.2 percent above the
efficient production costs in the quantity support auctions, where as it was on average
almost 20 percent above the efficient production costs in the price support auctions.
One plausible explanation is that the QSM helps the bidders find profitable
combinations so that the auction continues longer and either the efficiency of the bids
improves, or the mark-ups on bids are driven down (the required decrease in the
buyer’s total cost is 2% with every new bid), or both.
Paired, one-tailed t-tests were conducted to test the statistical significance of the
difference in the total cost to the buyer in quantity support and price support auctions.
Results of the t-tests for all the 48 designs indicate that the difference in the total cost to
the buyer (in favor of the quantity support auctions) is statistically significant (all p-
values smaller than 10-12).
120
There does not seem to be large variations in the total cost to the buyer in the quantity
support auctions. The only larger differences are between auctions with 15 vs. 30
bidders and auctions with equal vs. unequal bidder capacities. It is natural, that the
auctions with more bidders result in a lower total cost. The more there are bidders, the
more likely it is that there are bidders with low costs, and the average cost of the
winners should be lower. Also, the more there are bidders, the more there is
competition and the total cost to the buyer is driven closer to the production costs of
the bidders. One explanation for the better performance of auctions with equal bidder
capacities is that it increases the competition among the bidders, when all the bidders
can place the same bids. In that case, only cost matters. With unequal capacities a low
cost producer can be at a disadvantage due to a smaller capacity compared to another
bidder. Also, in the case of equal capacity case the bid quantities of each item (300 or
0) sum easily up to the total demand (600). Other reasons for the difference between
auctions with equal and unequal bidder capacities are discussed in section 8.2.4.6.
8.2.4.2 Efficiency of the Final Allocation
In the previous section I concluded that the total cost to the buyer is much lower in the
quantity support auctions than in the price support auctions. However, because the bid
prices also have mark-ups above the bidders’ production costs in them, we cannot
estimate the improvement in the efficiency of the final allocation. Because of larger
mark-ups, the final cost to the buyer in auction A could be higher than in auction B,
even though the allocation in A is more efficient. Thus, we need to “clean” out the
mark-ups in the bids. This is done by using the bidders’ cost to produce the bids. We
defined efficiency to be the ratio of the winning bidders’ combined cost of producing
the winning allocation and the total cost of the efficient (lowest total cost) allocation.
Let *
QI denote the combination of winning bidders, and *
,QiQ the combination of items
in bidder i’s winning bid in a quantity support auction, and *
,Qikq the quantity of item k
in that bid (notice that *
,Qikq = 0 for all items not in *
,QiQ ). The efficiency indicator can
then be expressed as
121
e
Ii
K
k
QikikQi
QC
qcF
efficiencyQ
Qi∑ ∑∈ =
+
=* 1
*
,, *
(27)
where Ce indicates the efficient allocation. Efficiency indicator for a price support
auction (efficiencyP) is defined similarly. Only in that case the combination of winning
bidders *
PI and the combinations and quantities of items *
,PiQ and *
,Pikq are taken from
the winning allocation of the price support auctions. If the final allocation does not
coincide with the efficient allocation, the ratio will be grater than one. For example, an
efficiency of 1.026 means that the cost of producing the winning allocation is 2.6%
above the efficient production cost. When efficiency improves, the ratio approaches
unity.
Table 17 presents the mean efficiency indicators of the auctions with quantity support,
and the benchmark case (price support). Again, the data have been grouped according
to the design parameters in order to study the effects of the chosen values.
Table 17 Efficiency ratios of final allocations in quantity and price support auctions
Quantity Support Price Support Mean st.dev mean st.dev
15 1.028 0.030 1.096 0.064 # of bidders
30 1.025 0.025 1.065 0.039 Normal 1.026 0.028 1.079 0.055
Econ. Of scope Large 1.026 0.028 1.082 0.056 Bid1 1.027 0.028 1.065 0.048 Bid2 1.025 0.026 1.067 0.042 Initial bids
Bid3 1.027 0.029 1.109 0.062 Equal 1.008 0.010 1.051 0.031
Capacities unequal 1.045 0.028 1.110 0.059 Full 1.025 0.027 1.080 0.054
Shortlist express 1.027 0.029 1.080 0.056
ALL 1.026 0.028 1.080 0.055
As can be seen from the figures in Table 17, the auctions where quantity support was
available for the bidders resulted in more efficient outcomes than the ones where only
price support was available. The cost of producing the winning allocation was on
average 2.6 percent higher than in the efficient allocation when quantity support was
used, and 8 percent higher, when only price support was used. Thus, part of the
122
difference in the final cost to the buyer in quantity support auctions vs. price support
auctions depicted in Table 16 can, in fact, be attributed to the higher efficiency of the
winning bidders in the quantity support auctions. Paired, one-tailed t-tests show that
the difference in the efficiency of quantity support and price support auctions is
statistically significant for all 48 designs (p-values all below 0.001, and vast majority
even below 10-10).
Notice also that the difference between the average efficiencies (Table 17) of quantity
and price support auctions is much smaller than the difference in the total cost to the
buyer (Table 16). This indicates larger profits for bidders (and a higher cost to the
buyer) in price support auctions. Based on these two tables we could also estimate the
size of the average mark-ups in the winning bids to be around 4% in the quantity
support auctions and 11% in the price support auctions. The higher mark-ups together
with the less efficient final outcomes caused the buyer to pay almost 20% extra over the
efficient production costs when only price support is available. When quantity support
is available, the buyer paid only a little over 6% extra in our experimental setting.
Looking at the efficiency ratios of the quantity support auctions, it can be seen that the
different design parameter values have very little effect on the efficiency of the final
allocation, except for the bidders’ capacities. When the bidders are symmetric in their
production capacities (and the capacities conveniently sum up to the total demand),
quantity support helps the bidders get very close to the efficient allocation. T-tests
suggest that the differences in the final efficiency between auctions with equal and
unequal bidder capacities are statistically significant. The variance of the production
costs of the winning allocation was also the smallest with equal production capacities.
This means that the teaming-up happens more easily when the bids are more similar.
The practical implication of this intuitive result is that it improves the efficiency of a
multi-unit auction to restrict the quantities that can be bid on to only a few levels. The
fact that there was very little difference in the efficiency of the final allocation between
the two types of shortlists used, and also very little difference in the total cost to the
buyer indicates that the “express” shortlist could be used. This would decrease the
bidders’ evaluation efforts when using the QSM and not worsen the buyer’s situation
significantly.
123
8.2.4.3 Total Cost to the Buyer in Non-Combinatorial Auctions
The total cost to the buyer in the quantity support auctions was also compared to the
total cost to the buyer in non-combinatorial auctions in which the items are auctioned
off individually (see section 8.2.3). Because of the assumptions we made about the non-
combinatorial auctions, the total cost to the buyer in those auctions coincides with the
lowest production costs for the single items. We wanted to compare the total costs to
the buyer (presented in Table 16) because we wanted to demonstrate that even though
allocative efficiency was not reached in the quantity support auctions, they were still
more profitable to the buyer than auctioning the items individually – even if assuming
that the efficient allocations of the individual item auctions could be reached with
competitive (= no excess profit) prices. The ratios of the total cost to the buyer in
quantity support auctions to the total cost in non-combinatorial auctions presented in
Table 18 verify this argument. Again, the data have been grouped according to the
design parameters.
Table 18 The ratio of total cost to the buyer in a quantity support auction and the lowest total cost in a non-combinatorial auction
Mean St.dev 15 0.910 0.069
# of bidders 30 0.873 0.052 Normal 0.897 0.064
Econ. of scope Large 0.886 0.064 Bid1 0.905 0.067 Bid2 0.885 0.063 Initial bids
Bid3 0.884 0.059 Equal 0.844 0.032
Capacities Unequal 0.939 0.052 Full 0.889 0.059
Shortlist Express 0.894 0.068
ALL 0.892 0.064
The total cost to the buyer is consistently around 10 percent lower in combinatorial
auctions than in single-item auctions. Naturally, the advantage of combinatorial
auctions is increased when the economies of scope are larger. Bidders can express their
synergies in combinatorial auctions, but not in single-item auctions. However, if the
non-combinatorial auctions are held as simultaneous auctions similar to the FCC
auctions described in section 3.1.2.2, and the bidders had better chances of winning
124
favorable combinations, the outcome of the benchmark case might not be as
inefficient as portrayed in this study.
8.2.4.4 Convergence Speed
In this simulation we also studied the convergence speed of the auction. The
convergence speed is an indicator of how much effort the bidders need to put into the
auction process. In our simulations an iteration is defined as a new bidder entering the
set of provisional winners. And because after the initial bids, the bids placed in the
auctions are all suggestions from the support mechanisms, moving from one iteration
to the next implies that the bidder found at least one bid suggestion profitable. The
total cost to the buyer was required to decrease by two percent from iteration to
iteration. Thus, the lower the final total cost – and usually also the better the efficiency
– the more there would be iterations in the auction. Thus, in our simulation setup it is
preferable to have the auction go on for as many iterations as possible – although at the
cost of bidder effort increasing.
We have already concluded in section 8.2.4.1 that the total cost to the buyer was on
average much lower in the quantity support auctions. From that directly follows that on
average the auctions with quantity support went on for more iterations than the price
support auctions, regardless of the design. The average number of iterations in all
quantity support auctions was 11.2 whereas in price support auctions the number of
iterations was 5.4.
However, the number of iterations only records the times the use of the support
mechanisms resulted in a new bid. The time and effort put into evaluating bid
suggestions which turned out to be unprofitable is not measured at all. One iteration
can take a long time if bidder after bidder uses the support mechanism to no avail
before finally one bidder finds a profitable suggestion. Thus, a better measure for the
convergence speed of the auctions would be the average number of times the support
mechanisms (price or quantity depending on the auction) was used in each iteration.
The number of bidders in the auction naturally affects the number of times a support
mechanism is used, because there are more inactive bidders to go through. Thus, the
number of times support was used per iteration is considered separately for the 15 and
125
30 bidder auctions. With 15 bidders, the average number of times support was used per
iteration was 5.3 for price support auction and 2.9 for quantity support auction. The
average total effort (number of times support was used during the entire auction) of the
bidders was then 28.62 in price support auctions and 32.48 in quantity support
auctions. With 30 bidders support was used per iteration 11.7 times in the price support
auctions, and 5.7 times in quantity support auction. Total effort of bidders was 63.18 in
price support auctions and 63.84 in quantity support auctions. Thus, even though the
price support auctions last for fewer iterations, the amount of effort the bidders have to
put in during the bidding process is almost as high as in the quantity support auctions,
and each iteration there are more futile attempts to find bids.
8.2.4.5 Usefulness of the Shortlist
We included the shortlist in the QSM because we believed that the original solution of
the QSP might not be the most profitable bid – at least not to all bidders. Thus it is also
interesting to study, whether the additional shortlist items proved to be useful or not.
In the simulations, whenever a bidder placed a bid suggested by the QSM, we recorded
whether the bid was the original (first) solution of the QSP, or whether it was one of
the additional bid suggestions on the shortlist. In the auctions where the bidders’
capacities were equal, the bidders chose the original shortlist item 94% of the time.
Thus, in those auctions, having the additional suggestions did not improve the
performance of the QSM that much. However, in the unequal capacities case the
original shortlist item was chosen over the other items only 57% of the time. Thus, we
concluded that having the additional shortlist items improved the performance of the
QSM, especially in the unequal capacities case.
8.2.4.6 Further Observations from the Quantity Support Auctions
Taking a closer look at the progress of simulated quantity support auctions, some
further observations can be made on how the auctions progress and what factors affect
the final outcome. In sections 8.2.4.1 and 8.2.4.2 I already demonstrated that one
simulation design parameter – bidders’ capacities - clearly affected both the efficiency
of the final allocation and the total cost to the buyer. However, looking into the
126
quantity support auction more closely allowed me to make more detailed observations
on what affects the progress and outcome of the auctions.
In addition, in the previous sections I have only considered the final outcome of the
auctions – be it the efficiency of the final allocation or the total cost to the buyer.
Another interesting viewpoint would be to study the change in the efficiency of the
allocation from the initial bid stream to the final (winning ) allocation.
Factors Affecting the Efficiency of the Auction Outcome
In Table 17 (section 8.2.4.2) I used average efficiencies to compare the outcomes of
different simulation designs. However, Parkes (2001) actually argues that the frequency
of exactly efficient outcomes is a better measure of the efficiency than e.g. average
outcome. Thus, I decided to study the outcomes of the quantity support auctions more
carefully. I observed that almost half of the auctions (1,136 out of 2,400) ended in an
efficient allocation (417/2,400), or a “pseudo efficient” allocation (719/2,400). With
“pseudo efficient” I mean two different situations in which the final allocation is not
efficient (the winners are not the efficient bidders or the winning bids are not the
efficient combinations), but should still be regarded as such in the analysis. Firstly, in a
situation in which the total cost to the buyer is already within 2% (the required
decrease in the total cost) of the efficient allocation production costs, it is not possible
for the efficient bidder to place the efficient bid without incurring a loss. Secondly, in a
situation in which the total cost is not within 2% of the efficient cost but in which
existing bids have such high mark-ups in their bids, it would require the incoming
efficient bidder to decrease her bid price below production costs to decrease the total
cost to the buyer by 2% (this is the threshold problem). These both situations are
considered “pseudo efficient” in the sense that the failure to reach the efficient
allocation is not the fault of the QSM. In fact, in almost all the pseudo efficient cases,
the QSM proposes the efficient bid to the efficient bidder, but the price required to
make the bid active is too low.
Of course, the definition of pseudo efficiency is tied to the bid decrement δ, which was
fixed at 2% in our simulations. However, because of that, as δ changes, the definition
for pseudo efficiency changes. For example, choosing a smaller δ, say 1%, would
127
narrow down the definition of pseudo efficiency leaving some of the current pseudo
efficient outside the scope. However, a smaller δ could potentially allow the auctions to
proceed one step further, making the final allocations again efficient or pseudo
efficient. Thus I do not expect different values of δ to change our results, because these
two effects may cancel each other out.
A common factor in the efficient and pseudo efficient cases is that the efficient
allocation usually consists of two bids only (they split the lot). Two bids is the
minimum number required, as the capacity constraints were adjusted so that it was not
possible for anyone to bid for the whole lot. This is the case in 1,068 out of the 1,136
efficient and pseudo efficient auctions. In the rest of the efficient or pseudo efficient
simulations, there are three bids in the efficient allocation, except for one simulation in
which there were four bids. Interestingly, in many of the pseudo efficient cases in
which the efficient allocation would have consisted of three bids, the actual winning
allocation still consists of only two bids. The production cost of the winning allocation
is so close to the efficient one that it was possible to bring the total cost to the buyer
within 2% of the efficient production cost.
The vast majority of auctions, in which the efficient allocation consisted of two bids,
were auctions in which the bidders’ capacities were equal. In the unequal capacities
case many bidders had capacities smaller than 300, so more than two bids were
required to fulfill the demand. Thus, based on this it would appear that the reason why
auctions with equal capacities led to significantly better efficiency than auctions with
unequal capacities (see Table 17) is the fact that in those cases the efficient allocation
consisted of only two bids. Also, when the capacities are equal (and sum up to total
demand), the solution space is in fact rather small. Thus, the efficient allocation is
easier to find, and it was found in 1,064 of the 1,200 simulations, while in the unequal
capacities case the efficient allocation was found only in 72 cases out of 1,200. Figure 1
below contrasts the cumulative distribution of the efficiency of auctions in which
capacities were equal to the efficiency of auctions with unequal capacities.
128
0
10
20
30
40
50
60
70
80
90
100
1 1.01 1.02 1.03 1.04 1.05 1.06 1.07 1.08 1.09 1.1 1.2Efficiency
% (cumulative)
Equal
Capacities
Unequal
Capacities
Figure 1 Cumulative distributions of the efficiency (compared to the production cost of the
efficient allocation) of auctions with equal capacities and unequal capacities
By looking at the distributions, it is clear that the efficiency of auctions with equal
capacities is generally better than in auctions with unequal capacities. About 90% of
auctions with equal capacities have efficiency ratios below 1.02, whereas the same is
true for only 20% of auctions with unequal capacities.
The assumption of equal capacities is analogous to a combinatorial auction in which
the demand is two units for each item, and bidders can only bid for one unit. This
comparison is straightforward, because all quantities in bids were either 0 or 300, and
total demand was 600. The single-unit combinatorial auction considered widely in
literature is even simpler, and thus I anticipate that the quantity support mechanism
would work even better in such auctions. The addition of multiple levels for capacities
increases the solution space tremendously. The initial bid stream (which has the same
number of bids as there are bidders in the auction) now covers only a very small
portion of the solution space. The QSP always searches the complements for the new
bid from the bid stream, and hence the content of the initial bid stream becomes very
important. In our simulations, when the efficient allocation consisted of more than two
bids, it was necessary that at least one of the efficient bids be in the initial bid stream.
However, having one or more efficient bids in the initial bid stream was not sufficient;
the final allocation of the auction might still not be efficient. When the efficient
129
allocation consisted of two bids, however, the efficient bids needed not be present in
the initial bid stream; the efficient allocation was still found almost every time.
The total cost to the buyer behaved similarly as the efficiency of the final allocation.
Naturally, the efficiency of the final allocation is correlated with the final cost to the
buyer, because efficient bidders can afford to drive the price lower. However, an
efficient allocation does not guarantee a low total cost to the buyer (12% of efficient
final allocations resulted in a total cost more than 5% above the efficient production
cost). It would appear that the best indicator is again the capacities of the bidders (or
the number of bids in the efficient allocation, which is closely related to the capacities
of the bidders). Figure 2 depicts the cumulative distributions of the total cost to the
buyer in auctions with equal and unequal capacities.
0
10
20
30
40
50
60
70
80
90
100
1 1.01 1.02 1.03 1.04 1.05 1.06 1.07 1.08 1.09 1.1 1.2 1.3Cost to Buyer
% (cumulative)
Equal
capacities
Unequal
capacities
Figure 2 Cumulative distributions of buyer's total cost (compared to the production cost of the
efficient allocation) in auctions with equal capacities and unequal capacities
From Figure 2 it is clear that the auctions with equal capacities led to lower total cost
to the buyer than auctions with unequal capacities. Almost 90% of auctions with equal
capacities ended up in a total cost to the buyer with 4% of the efficient production cost,
whereas less than 5% of auctions with unequal capacities achieved the same level.
Final Efficiency vs. Initial Efficiency
Thus far only the efficiency of the final allocation has been studied. However, a valid
question is, what the efficiency of the auction was before the bidders began using the
QSM. This is especially valid since the set of initial bids was determined in the
130
simulation design. Therefore, it is a possibility that certain auctions end in a more
efficient outcome simply because they started off with a more efficient allocation. Also,
it would be interesting to know, if the QSM improves the efficiency of certain types of
auctions more than some others. According to Table 17, the way the initial bid stream
was constructed did not have a very big effect on the efficiency of the final allocation.
However, from Table 17 one cannot study possible effects two or more design
parameters could jointly have.
Thus, I have calculated the difference between the initial efficiency (= the production
cost of the provisional winners in the initial bid stream) and final efficiency (= the
production cost of the winning bidders at the end of the auction). The average
improvement in efficiency, i.e. the decrease in the production costs, is reported in
Table 19 as a percentage of the efficient production cost. The data in Table 19 is
aggregated over all simulations according to the simulation design parameters.
Table 19 Average improvement in efficiency in the quantity support auctions
Average
improvement (as % of eff. prod.cost)
equal 3.6
unequal 9.0
15 8.5
30 4.1
normal 6.2
large 6.4
express 6.2
full 6.3
bid1 5.5
bid2 4.7
bid3 8.6
Shortlist
Initial bid
Design variable
Capacities
Bidders
Econ. of scope
The results in Table 19 indicate that the improvement in efficiency during the auction
is larger if the bidders’ capacities are unequal, there are only 15 bidders, or if the logic
of “Bid3” is used to create the initial bid stream. Studying the interactions of these
three design parameters, Table 20 was constructed. In the table, improvements in the
efficiency were averaged over the three interesting design parameters. The size of the
economies of scope and the compilation of the shortlist do not seem to have an effect,
so they were not studied further.
131
Table 20 Average improvement in efficiency (as % of efficient cost) broken down according to three design parameters
Equal Unequal Equal Unequal
Bid1 3.3 12.1 1.6 4.8
Bid2 2.9 9.5 1.9 4.5
Bid3 7.5 15.3 4.1 7.5
30 bidders
Initial
bids
15 bidders
It is clear from Table 20 that the efficiency improved the most in the auctions in which
there were 15 bidders and the bidders’ capacities were unequal. Also if the logic Bid3
was used to construct the initial bid stream, the improvement in efficiency was larger
than if other logics were used.
In order to understand what is behind the figures in Table 20, I constructed a table
from the average final efficiency ratios in Table 17 , but regrouped them according to
the grouping in Table 20.
Table 21 Average final efficiency ratios broken down according to three design parameters
Equal Unequal Equal Unequal
Bid1 1.009 1.049 1.009 1.042
Bid2 1.009 1.046 1.008 1.039Bid3 1.006 1.050 1.008 1.043
Initial
bids
15 bidders 30 bidders
As can be seen from Table 21, the differences in final efficiency are not dependent on
the number of bidders or the logic by which the initial bids are constructed. The only
design parameter explaining the differences is the bidders’ capacities – the conclusion
to which we have already come in the previous sections.
Notice that there are large differences in how much the efficiency improved during the
auctions, but that there is much less variation in the final efficiencies of the auctions.
This means that there must be large differences in the initial efficiencies of the
auctions. Indeed, the average initial efficiency is worse (i.e. efficiency ratio is higher)
for auctions in which there are 15 bidders and the bidders’ capacities are unequal, or
when the initial bids are created with Bid3. What is very interesting is that the QSM
seems powerful enough to smooth out the initial differences caused by the smaller
number of bidders, and the bid creation logic Bid3. However, it would appear that the
QSM it cannot quite tackle the inefficiency caused by the unequal bidders’ capacities.
132
9 IDEAS TO IMPROVE THE QSM18
The results of the second simulation study indicated that there is still room for
improvement in the QSM. Whenever there were more than two bids in the efficient
allocation (which corresponds to the case of unequal bidder capacities in our
simulation setting), it became difficult for the QSM to lead the auction to the efficient
allocation. The contribution of this chapter is to discuss improvement ideas to the
QSM and their applicability.
9.1 Varying the Bid Decrement
The bid decrement δ refers to the amount by which the total cost to the buyer should
decrease every time a new bid enters the set of provisional winners. In both simulation
studies presented in Chapter 8 the decrement was fixed, although the size of the
decrement was different (5% of current total cost in the first simulation and 2% in the
second). In section 8.2.1 I discussed the effects the size of the decrement can have on
the convergence of the auctions. However, so far I have not discussed the effects the
fact that the decrement is fixed has on the QSM and thereby on the convergence and
outcome of the auction.
In its original form, the QSP forces the incoming bid to bear the whole burden of
decreasing the total cost to the buyer by the fixed decrement δ (defined as a percentage
of current lowest total cost). Especially, when the bid quantities are small, and
therefore also the price attached to the bid is low, the burden is unreasonable. The
price for that one bid has to go very low in order for the total cost – which possibly
consists of several other bids – to decrease by a fixed decrement. This leads easily to
unacceptable bid suggestions from the QSP and a definite bias towards large bid
suggestions. Choosing a small δ would make it somewhat easier for small bids to be
acceptable, but a small δ slows down the convergence of the auction, and it would not
change the fact that large bids would still be favored.
18 The ideas presented in this chapter are based on a brainstorming session with Professors Hannele Wallenius, Jyrki Wallenius and Murat Köksalan.
133
Thus, we developed the idea to make the price decrement “dynamic”, i.e. dependent
on the bid size. Following the idea presented by auction scholars in relation to Federal
Communications Commission (FCC, 2000) auction # 31 (a combinatorial auction),
we thought it might be a good idea to have the decrement depend on the items and
item quantities in the bid. Only if the incoming bid is for the whole demand should
the decrement be effective fully. For example, assume the full decrement is set at ∆ =
5%, then, if the bid is for only “half” of the whole bundle, the required decrease would
be only 2.5%.
The formulation of such dependence in terms of items and item quantities in a
combinatorial auction is difficult. The items might be very different from each other,
so that for example half of the total cost could be accrued from one single item, and
the rest from all other items. Therefore the definition of “a half bundle” is not trivial,
unless the item quantities were 50% of the total demand for each item. The auction
owner (buyer) could assign relative weights to the items, so that the size of each bundle
in monetary terms could be evaluated. Also, instead of buyer defined weights, the dual
prices could be used as relative weights to estimate the value (monetary size) of any
given bundle. The total cost of the bid could be estimated using the dual prices, and
the cost of the bid would then be compared to the total demand, to see how large of a
portion the bid represents. A potential problem can arise if some dual price is zero,
because that can underestimate the costs drastically, and allow for too small a delta.
Even a simpler idea is to approximate the “size” of the bid by comparing the bid price
in the new bid to the total lowest cost to the buyer. Let ∆ denote the decrement by
which a bid for the whole demand is required to lower the total cost, and let δ denote
the decrement by which an incoming bid should decrease the total cost. Then δ would
be defined as
∆
+
=
∑∑= =
new
N
i
n
j
ijij
new
pxp
p
i
1 1
δ (28)
where pnew is the price of the incoming bid, pij the price of bidder i’s jth bid in the bid
stream, and xij { }1,0∈ indicates which bids are among the current winners. However,
134
defining δ in this manner results in a nonlinear constraint in the QSP. Thus, we
decided to approximate the new total cost in the denominator with the old total cost
C* (the solution of the WDP without the incoming bid), which is easily available.
Because C* is by definition an upper bound for the denominator, it results in a slightly
smaller δ than would be appropriate. However, we do not believe the difference would
be significant.
The constraint (7) in the QSP (the one requiring that the total cost to the buyer
decreases) would then take the form of
( ) *1
**
1
1 1
1 1
Cpxp
or
CC
ppxp
new
N
i
n
j
ijij
new
new
N
i
n
j
ijij
i
i
≤∆++
∆−≤+
∑∑
∑∑
= =
= =
(29)
When testing the dynamic delta as suggested above in (29), it was discovered that the
use of a dynamic delta led to the QSM to suggest drastically smaller bids, and the
efficiency of the final allocation deteriorated. There are two reasons for this. First, the
original QSM using dual prices already has a bias towards small bids, as the linearized
cost function underestimates the cost of small bids and conversely overestimates the
cost of large bids. The dual prices are consistently above the true variable cost
parameters and there is no fixed cost element. Thus, small bids appear cheaper than
they should, and large bids more expensive. This bias is aggravated with the dynamic
delta, which requires less reduction in the total cost from small bids.
Thus, the fact that a fixed delta introduces a bias towards larger bids conveniently
counteracts the bias towards small bids inherent in the QSM. That is why the original
QSM worked as well as it did. What was puzzling though, was that when the
approximated (linear) cost function in the QSM was replaced by the bidder’s true cost
function, the dynamic delta did not perform better than the fixed delta – in fact, it did
worse. Naturally, the relative sizes of the fixed and dynamic deltas, affect the results.
We tested two different levels for the variable delta, and the results were the same. We
135
did not pursue the research into the dynamic delta because it seemed clear that it
would not significantly improve the efficiency of the final allocation.
9.2 Initial Bids in the Form of Ranges
One of the reasons the efficient allocation was not found in many of the simulated
auctions was that the initial bid stream did not contain the bids from the efficient
allocation, and therefore also the QSM cannot find the missing efficient bids. Thus,
one way to try to improve the performance of the QSM would be to increase the initial
bid stream so that it would encompass more combinations. One idea would be to
express the initial bids in the form of ranges. The bidders would give lower and upper
bounds on the item quantities. Because the bids are linear within the bounds, also the
price needs to be expressed linearly. Providing a per-unit price for each item is
somewhat against the idea of combinatorial auctions, but since the price is only valid
within the quantity ranges, it is only a minor compromise. The subsequent bids and
the bid suggestions of the QSM would still be fixed (without ranges) and with only one
price for the whole package, just as before. However, with the initial bid quantities
expressed as ranges and per-unit prices the WDP needs to be adjusted from the original
formulation (Eq. (2)). Now the qijk’s are also variables to be determined in the solution,
hence the objective function takes the form
∑∑∑∑= == =
+N
i
n
j
ijij
N
i
K
k
kikii
i
pxqpx1 21 1
111min (30)
where qi1k is the quantity for item k in bidder i’s initial (j = 1) bid, pi1k is the per-unit
price of item k in bidder i’s initial bid, and pij is the bundle price for bidder i’s
subsequent bids. The following constraints need to be added so that the ranges
specified by the bidders in their initial bids are taken into consideration when the qi1k’s
are solved for:
111 iikkiiik xUqxL ≤≤ (31)
Where Lik is the lower bound and Uik the upper bound for the quantity of item k in
bidder i’s initial bid. Similar adjustments need to be made to the QSP as well.
136
In the initial tests we ran, this modification of having bidders place the initial bids in
the form of ranges did not have much of an effect on the efficiency of the final
outcome. What happened was that the non-zero bid quantities were chosen to be
either at the upper or lower bound. The bid space covered by the initial bid stream did
expand, but only limitedly. Towards the end of the auction, the initial bids are edged
out from the winning allocation. The bid quantities in the bids suggested by the QSM
did change as a result of the new initial bids compared to the original situation in
which the initial bids were formulated as bundle bids. However, the WDP’s and the
QSP’s repeatedly chose the same upper or lower bound values for the initial item
quantities. Hence, the effect of the modification remained modest. Of course, our
findings are affected by our experiment set-up, hence it is possible that in another
setting the results could be better.
9.3 Allowing for Shortages and Excesses in the Supply
To make the QSM more flexible in finding good bid suggestions it might make sense
to allow for a shortage in the supply of some of the items, and conversely allow for
supply exceeding demand for some other items. This way the exact complements for
the bid suggestions do not have to exist in the bid stream.
The formulations of the WDP and QSM already allow for excess supply, but it is
hardly ever present because having extra units of some items increases bidders’ costs
and thereby also the total cost to the buyer making such solutions rarely optimal.
However, in reality, such solutions might not be all that undesirable for the buyer. In
real life, procurement situations are rarely one time events. Rather, companies
purchase the same items over an over again. Thus, in order to truly allow for the WDP
or QSM to find solutions in which there is excess supply, some compensation should
be given from the excess units. This compensation reflects the fact that now the buyer
is in fact getting more for her money, and it makes the comparison with smaller
packages more equitable.
The original formulations of the WDP and QSM do not allow shortages in any items.
However, relaxing this constraint would give more flexibility in the solution of the
problem than simply allowing excess supply. The shortages should be penalized,
137
though, to indicate that it is not desirable to have a shortage but that it could be
tolerable if the overall deal is then better. Allowing for shortages makes sense in
practice, because one auction is hardly the only opportunity to procure these items.
The missing units of the items can most likely be procured in some other auctions or
then directly from the market. Naturally, if these conditions do not apply for some of
the items, excesses and shortages can be allow for only a subset of the items in the
auction.
Denote the excess in the supply of item k with ek and the shortage with sk. The per-unit
penalty (i.e. extra cost) from shortage of item k is denoted with Sk and the
compensation from excess supply with Ek. The WDP now takes the form
{ }1,0
,...,1..
)(min
1 1
11 1
∈
=∀−+=
−−
∑∑
∑∑∑
= =
== =
ij
kkk
N
i
n
j
ijkij
K
k
kkkk
N
i
n
j
ijij
x
Kksedqxts
sSeEpx
i
i
(32)
and similar additions need to be made to the QSP as well.
When testing the new formulation we soon realized that it is sensitive to the choice of
Ek and Sk. If they were not chosen correctly, there would never be excesses or shortages.
In addition, we noticed that the QSM very quickly clears away all the shortages that
may have appeared in the initial stages of the auction. It cannot be a profit maximizing
solution for any bidder to leave some shortage in the supply whenever the penalty from
shortage Sk is larger than the bidder’s variable cost ck for item k (unless restricted by a
capacity constraint). And, if Sk < ck, this effectively means that the buyer can acquire
the item cheaper somewhere else. Hence there would be no point in buying anything
from this bidder, and the auction would become useless.
Based on our findings we concluded that the excesses and shortages formulation would
not help in improving the efficiency of the final allocations in the quantity support
auctions discussed in the previous chapter. However, we think that the excesses and
shortages formulation could prove useful in one-shot auctions. By one-shot auctions we
mean all non-iterative auctions, in which the WDP is solved only once during the
138
entire auction. This means that the bidders cannot receive any feedback on how to
improve upon their bids, and it is possible that the bids placed in the auction are not
good complements to each other. By allowing the possibility of excesses and shortages
in the formulation of the WDP, the buyer could potentially find different and perhaps
more efficient solutions.
139
V THE GROUP SUPPORT MECHANISM
After testing the improvement ideas in Chapter 9 we concluded that minor changes in
the QSM most likely would not improve the efficiency of the final allocations. The
further observations from the results of the second simulation study (section 8.2.4.6)
gave indication that one problem with the QSM is that it optimizes only one incoming
bid at a time. First of all, when you can adjust only one bid at a time, it is not possible
to overcome the threshold problem. As was explained already earlier, the threshold
problem arises when the bidders should jointly revise prices in their existing bids so
that together (possibly with one new bid) they could beat the current winner(s). Thus,
by giving support to only one bidder at a time, the threshold problem remains. Also, it
would appear that the “puzzle problem” is broader than we anticipated. The QSM was
designed to alleviate the puzzle problem, in other words, to solve for the shape of the
“last missing piece to the puzzle”. Undoubtedly, in this capacity the QSM was
successful: it was able to find the “missing pieces” allowing the quantity support
auctions to continue much further than the price support auctions. However,
oftentimes, if striving for the efficient allocation, it is not enough to solve only for the
last missing piece. When the other bids that form the efficient allocation are not in the
bid stream, the QSM cannot find the last missing efficient allocation bid. In fact, in
this case there are several pieces missing from the puzzle, and the shape of just one of
them cannot be solved. In order to significantly improve the allocative efficiency, the
QSM should be modified to address this broader puzzle problem as well as the
threshold problem.
10 THE GROUP SUPPORT MECHANISM19
The contribution of this chapter is to introduce another bidder decision support tool
we have designed for a semi-sealed-bid multi-unit combinatorial auction. I also present
a detailed example auction to explicate the use of the GSM. The design follows the
logic of the QSM, but has significant improvements. Thus, one contribution is that we
19 Material presented in this chapter will be published in Köksalan, Leskelä, Wallenius and Wallenius (2009), forthcoming in Decision Support Systems.
140
have designed a way to use information in bidders’ bids to improve the approximation
of the bidders’ cost functions.
We call the new tool the Group Support Mechanism (GSM). The purpose of the
GSM is to circumvent the problems arising from optimizing only one incoming bid at
a time, which make it difficult for the QSM to lead auctions to an efficient allocation.
Following the logic of the QSM, the GSM looks for bids that would become
provisional winners. The difference is that the GSM suggests a combination of bids
that either together satisfy the entire demand, or team up with one or more of the
existing bids in the bid stream to become active as a group. This should improve the
efficiency of the final allocation, because the GSM also chooses how many new bids it
suggests, and is therefore not as dependent on the existing bids as the QSM. The QSM
is a special case of the GSM, because the GSM will also support only a single bidder
when finding it the optimal course of action. However, the GSM is free to suggest any
number of incoming bids at a time, and the added flexibility should result in more
efficient outcomes than the use of QSM.
10.1 The Group Support Problem
At the heart of the GSM is the Group Support Problem (GSP). The formulation of the
GSP is related to the QSP – which is only natural since the underlying logic behind
the two mechanisms is similar. The QSP was designed to maximize the profit of the
incoming bidder. Conversely, the GSM attempts to maximize the joint profit of the
provisionally winning bidders, that is, the profits of the new bids and existing bids to be
included in the provisionally winning combination. The inclusion of the profit of
existing bids in the GSM may appear somewhat counterintuitive, because the purpose
of the QSM was to serve the incoming bidders and find new, profitable bids for them.
However, not including the profit of the existing bids would create a bias towards filling
the set of provisional winners with new bids. The existing bids would not be teamed-up
with, even if they were good matches. Thus, the profits of both old and new bids are
included in the objective function of the GSP.
The notation in the formulation of the GSP is similar to that used in the QSP: there
are K items, dk (k = 1, … , K) units of each item requested by the buyer, N bidders, and
141
ni bids from each bidder i. Again, xij’s indicate the statuses of the old bids, (.)~ic is the
approximated cost function of bidder i, and qi,new,k’s indicate the item quantities
(elements of Qi,new) and pi,new’s the prices in the bid suggestions to be solved in the GSP.
C* is the current lowest total cost to the buyer, δ the percentage by which the total cost
is required to decrease, and aik represent the bidders’ capacities. At any point in time
the bidders can be divided into two sets: active bidders, who are among the provisional
winners, and inactive bidders who are not. Let I denote the set of inactive bidders. The
GSP solves for a combination of new bids for the inactive bidders which, as a group
(and possibly together with some of the existing bids in the bid stream), become active:
(GSP) ( )
( )
( )
{ }
0,,,
, 1,0,
,,...,1
1
1
0
0
0~
,...,1
1
(~max
,,,
,
,,
,
1
1
,,
,
1
,,
,,
,,
1 1
*
,
1 1
1 1
≥
∀∈
∈∀=≤
∈∀≤+
∉∀≤
∈∀≥−
∈∀≥−
∈∀=+−−
=∀≥+
−≤+
+
−+−
∑
∑
∑
∑∑∑
∑∑∑
∑∑∑∑
=
=
=
∈= =
∈= =
∈= =∈
knewinewiii
newiij
ikknewi
newi
n
j
ij
n
j
ij
newinewi
newi
K
k
knewi
iinewiinewi
k
Ii
knewi
N
i
n
j
ijijk
Ii
newi
N
i
n
j
ijij
Ii
i
N
i
n
j
ijijiij
Ii
i
qpes
jixx
IiKkaq
Iixx
Iix
IipMx
Iixq
IiseQcp
Kkdqxq
Cpxps.t.
exQcps
i
i
i
i
i
δ
εε
(33)
(34)
(35)
(36)
(37)
(38)
(39)
(40)
(41)
The objective (33) is to maximize the combined (approximated) profit of all the
bidders by choosing the statuses of the old bids (xij), and the quantities (qi,new,k) and
prices (pi,new) in the new bids (i.e. sum of new profits ei defined in (36)) suggested for the
inactive bidders. The constraints (34), (35), (39), and (40) are the same as in the QSP.
Notice that since the objective function maximizes the total profit of the active bidders,
it does not discriminate against solutions in which some bidders accrue a loss. Because
142
the new bids in the combination suggested by the GSM will become active only if all
the bids are accepted by the bidders, such solutions are not desirable. However, we
acknowledge that the profits we use in the formulations are approximations, and if we
require that all profits be at least zero, we might not get a feasible solution even if one
existed. This could be the case if we have overestimated the bidders’ costs. Thus, we
allow losses, but add a penalty for losses (si) into the objective function and formulate
corresponding constraints (36). The addition of ε (a very small positive constant) in the
objective function implies that the primary objective is to avoid losses and profit
maximization is secondary. Essentially, the GSP chooses the most profitable
combination of bids from combinations that do not result in losses for any bidder, if
possible. If losses cannot be avoided, they will still be minimized. Constraints (37) and
(38) (where M is any large number) ensure that a price in a new bid can be positive
only if at least one item assumes a positive value and that the status variable for the new
bid, xi,new = 1, when the bid is not empty.
10.2 Customizing Cost Functions Approximations
The formulation of the GSP – just like the formulation of the QSP – requires an
approximation of the bidders’ cost functions. Thus, the second building block of the
GSM is the procedure through which the cost functions are approximated. In the
QSM, the cost function approximation was a linear cost function using dual prices as
per-unit cost parameters. The QSM used the same cost function estimate for all
bidders, and thereby the bid suggestions of QSM were anonymous (i.e. it gave the
same suggested bids regardless of the bidder). However, it does not make sense for the
GSM to be anonymous. If each bidder did not have a unique approximation of the cost
function, the optimization algorithm would assign the bid suggestions randomly to
some bidders, (as there would be numerable alternative optima) and it would pool the
quantities into one large bid or few large bids (as allowed by the capacity constraints).
We hold on to the assumption that bidders do not want to disclose any cost
information, and thus the only information we have on the bidders’ costs is information
from the bidders’ bids in the bid stream. We use the bid information to customize the
cost function approximations for each bidder.
143
To start off, the form of the cost function to be used in the approximation was chosen.
We decided to try out the same functional form as was used in the two simulation
studies in Chapter 8. This way we could compare the results of the GSM auctions to
the QSM simulations. To approximate the bidders’ cost functions we designed an
inverse optimization problem (in the spirit of Beil and Wein, 2003; see also Zionts and
Wallenius, 1976), which utilizes the information we get in the form of bids in the bid
stream. We assume that bidders do not place bids in which costs exceed the price.
Thus, the task of the inverse optimization problem is to find a set of cost function
parameters, which are consistent with a bidder’s bidding behavior (see (42)). Also, we
made some assumptions that allowed us to pose some constraints on the cost
parameters. First, we assumed that by including an additional item into the bundle
should not decrease the total cost, i.e. F12 ≥ F1 and F12 ≥ F2, (see constraints (45)).
Secondly, we assumed that there would be economies of scope between the items, i.e.
F12 ≤ F1 + F2, (see constraints (46)). Also, in order to constrain the feasible set a little
more, we assumed that upper and lower limits for all the cost parameters can be
derived for any particular industry (constraints (43) and (44)).
The constraints of the Cost Estimation Problem (CEP) are:
[ ]
,'
,...,1
,...,1)(~
'
Γ⊂∀≤
Γ⊆∀⊂∀≤
=∀≤≤
Γ⊆∀≤≤
=∀≥−
∑∈
LFF
LLLFF
Kkucl
LuFl
njQcp
LParL
iLiL
iLiL
ikikik
iLiLiL
iijiij
t
t
ε
(42)
(43)
(44)
(45)
(46)
where ε is a small positive scalar, Γ is the set of all possible item combinations, and L,
L’ are subsets of Γ, Par[L] refers to all possible partitions of L, and Lt is an element of
Par[L]. A partition of set L is the group of disjoint sets, which together form L.
These constraints (42) – (46), however, still leave a vast range of feasible options to
choose from. Therefore, the choice of the objective function to a large extent
144
determines the values for the cost parameters. Thus, the question becomes how to
choose the objective function.
The use of different objective functions will lead the CEP to choose different points in
the feasible set. Without any further information on the bidders’ cost functions besides
the constraints that make up the feasible set, there is no way of knowing, which point
would be better than some other point. In other words, it is impossible to say which
objective function would provide the best – or even good – approximations of the
bidders’ cost functions. Thus, we chose simply to maximize the sum of the cost
function parameters as the objective of the CEP, that is,
max1
ij
K
k
ik
L
iL pcF ∑∑=Γ⊆
+ (47)
and designed the following iterative scheme to approximate the bidders’ cost function
parameters and to narrow down the feasible set as the auction progresses.
First, lower and upper bound estimates for the cost function parameters are set. At the
beginning of the auction the bounds coincide with “industry estimates”, or in our
experiments, the ranges of the distributions. The objective function will drive all the
parameters to their upper bounds in the absence of any bid information to provide
contradicting evidence. This may naturally be an over estimation of the cost functions,
but it will not prohibit the GSP from finding bid suggestions, since losses are allowed
in the formulation. Thus, it can still suggest the bids to the bidders even though the
cost function approximations suggest that the bids would result in losses, and it is
possible that the bidders will actually find them profitable. If the bidders accept what
appeared to be unprofitable bids, it has the added benefit that now we get contradicting
information and can update the estimates for the cost function parameters. The
updated estimates are then set as the new upper bounds for the parameters, and the
auction continues. If we started from a lower bound estimate for the cost parameters,
and the GSP suggests bids in which all the estimated profits are positive, the
acceptance of the bids is expected and would not give us any new information on the
cost function parameters (the old estimates are still consistent with the new evidence).
Naturally, in this case, if the bidders declined the bid suggestions we would get new
145
information. However, it is desirable that the bidders accept the bid suggestions
because that way the auction progresses. In order to maximize the information
obtained from new bids and to speed up the auction process it makes more sense to
start from the upper bound estimates.
It is worth noting that the GSP will not offer bid suggestions to all inactive bidders
unless there is room for everybody in the set of provisional winners, which is more
unlikely the more there are participants in the auction. Thus, some bidders are not
suggested a bid by the GSP, and without any bid information the cost function
estimate would not be lowered, which decreases the probability that the bidder would
be offered a bid suggestion the next time. Some bidders could get stuck in this loop,
and never be suggested anything. We could consider adding a constraint requiring that
the bidder requesting support would be guaranteed a bid suggestion (not to upset the
bidder), but so far we have not added any such constraints. Instead, recognizing that
the estimates are above the true parameters, we decided to decrease the estimates by
1% for each bidder who is not suggested anything to improve their chances to receive a
suggestion in the next round. We chose the decrement to be 1% in order to make only
small adjustments in the estimates. Increasing the decrement could reduce the number
of iterations needed to get the GSM to suggest a bid for the bidder, but a smaller
decrement allows us to get closer to the true estimates. If, in the next round, the bidder
receives a bid suggestion and accepts it, the upper bounds are replaced by the 1% lower
estimates. If it turns out that the new estimate was too optimistic (the GSM offers a bid
it thinks is profitable, but the bidder declines), we solve the cost function estimation
problem again with the rejected bid added to the constraints, and receive an updated
estimate of the parameters.
10.3 Example of an Auction with GSM
In order to present, how the GSM actually works, we designed an example auction,
which is described below in detail. The example aims at clarifying how the GSP and
CEP are formulated and solved, and how the auction advance. The outcome of the
auction was also analyzed in order to evaluate the performance of the GSM.
146
10.3.1 Initiation Phase
In this example auction there are three items, 600 units of each item demanded by the
buyer, and 10 bidders. All bidders are assumed to be “glocal” meaning that they have
production capacity for all three items but they are willing to settle for any item and
unit combination as long as it is profitable. The bidders are assigned capacities.
Because the second simulation study demonstrated that the case of unequal capacities
was more challenging, the capacities in this example are unequal as well. The three
possible levels for the bidders’ capacities are the same as in the simulation study (150,
225 and 300). The bidders are also assigned cost functions. Assigning specific cost
functions allows for the evaluation of the profitability of the bid suggestions, and for the
solution of the efficient allocation, against which the auction outcome can be
evaluated. The cost functions are of the same for as in the simulation studies presented
in Chapter 8. The parameters are chosen randomly from uniform distributions, and for
the sake of simplicity no difference is made between the items: variable costs cik vary
from [53.3; 66.7], fixed costs FiL for single items from [1,777.8; 2,222.2], for pairs of
items from [2,666.7; 3,333.3], and for combinations of three items from [3,555.6;
4,444.4]. It is made sure that the cost functions exhibit economies of scope, and the
ranges for the parameters ensure that the fixed cost cannot decline as a result of an
additional item being included into the bundle.
Because the GSM cannot be used unless there is already a feasible solution to the
WDP (and thereby a total cost to the buyer), we assume that at the beginning each
bidder places one bid. The bidders place a bid on the combination for which their
fixed cost is the lowest compared to the expected value of the cost. The item quantities
are set at the upper bounds of the capacity constraints. The initial bid stream is
reproduced in Table 22.
147
Table 22 The initial bids and provisional winners
Bid Item 1 Item 2 Item 3 Price (€) Status*
x11 0 225 150 27,199 0
x21 150 0 150 22,954 1
x31 225 225 150 45,990 0
x41 0 225 0 17,613 0
x51 0 225 150 25,594 0
x61 225 150 300 49,356 0
x71 300 0 0 22,007 0
x81 0 225 225 33,210 1
x91 225 150 0 27,831 1
x10,1 225 225 225 50,076 1 * 1 = active, 0 = inactive; Total cost to buyer = 134,071€
The next step is to estimate the cost functions of the bidders. The CEP is unique for
each bidder, but only with respect to the constraints derived from the existing bids (42).
The constraints resulting from the economies of scope assumption are the same for
each bidder. Also initially, the upper and lower bounds for the parameters (which are
set to be the upper and lower bounds of the uniform distributions from which the
bidders’ true cost function parameters were drawn) are identical for all bidders. As the
auction progresses, the upper and lower bounds will differ across bidders. The
objective function, which was chosen to be the sum of the parameters, is the same
across bidders throughout the auction. For example, for Bidder 1 the CEP assumes the
following form. For the sake of simplifying the notation, the index for the bidder has
been left out.
148
123123123213123312
233213311221
231231312312123
323223
313113
212112
123
23
1
3223
321123231312321
, ,
, ,
, ,
,
,
,
3,...,17.663.53
4.44446.3555
3.33337.2666
3,2 3.33337.2666
3,2,1 2.22228.1777
15022527199..
max
FFFFFFFFF
FFFFFFFFF
FFFFFF
FFFF
FFFF
FFFF
ic
F
F
jF
iF
ccFts
cccFFFFFFF
i
j
i
≥+≥+≥+
≥+≥+≥+
≥≥≥
≥≥
≥≥
≥≥
=∀≤≤
≤≤
≤≤
=≤≤
=≤≤
≥−−−
+++++++++
ε
(48)
where ε is a small positive constant. Once the CEP has been solved for each bidder, we
have the following initial estimates for the cost function parameters:
Table 23 Initial estimates for the cost function parameters
Bidder 1 Bidder 2 Bidder 3 Bidder 4 Bidder 5 Bidder 6 Bidder 7 Bidder 8 Bidder 9 Bidder 10
F i 2222.2 2222.2 2222.2 2222.2 2222.2 2222.2 2222.2 2222.2 2222.2 2222.2
F ij 3333.3 3333.3 3333.3 3333.3 3333.3 3333.3 3333.3 3333.3 3333.3 3333.3
F 123 4444.4 4444.4 4444.4 4444.4 4444.4 4444.4 4444.4 4444.4 4444.4 4444.4
c 1 66.7 66.7 66.7 66.7 66.7 66.7 65.9 66.7 64.4 66.7
c 2 61.6 66.7 66.7 66.7 54.5 66.7 66.7 66.7 66.7 66.7
c 3 66.7 64.1 66.7 66.7 66.7 66.4 66.7 66.1 66.7 66.7
As can be seen from the table, the CEP chooses the upper bounds of the fixed cost
parameters for all bidders, and for most variable cost parameters, except for those
bidders, whose initial bid indicates that it is not possible that all the cost parameters
would be at the upper bound.
The auction is a continuous auction, that is, the WDP is solved after each incoming
bid. However, since in this example, bids only enter based on the suggestions of the
149
GSM, it is easier to present the example “round by round”, in which one round
continues as long as there is a change in the solution of the WDP (i.e., the set of active
bidders changes).
10.3.2 The Auction
Round 1
It is now time to solve the GSP for the first time. The decrement δ by which the total
cost to the buyer is required to decrease from round to round is set at 2%. The cost
function parameters in Table 23 are inserted in the cost function in the GSP (33).
However, because the cost function is discontinuous (the fixed cost term depends on
the combination of items in the bid), the formulation becomes somewhat more
complex, and it bears resemblance to the formulation of the QSP with the true cost
function (Appendix 2) and of the efficient allocation problem (Appendix 3). A set of
auxiliary variables yi,jkl is defined to construct constraints that guarantee that the correct
fixed cost is taken into consideration in the objective function (yi,jkl = 1 if the items in
combination j,k and l all have a non-zero value, yi,jkl = 0 otherwise). Also, for the same
reason we need to create new variables for the item quantities in the incoming bids:
one variable per item per combination it is in. If there are K items in the auction, each
item is in 2K-1 combinations, so in this case each item is in four combinations.
Thus, the cost function ( )Qc~ for bidder i to be inserted into (33) takes the form
( )
123,123,23,23,13,13,12,12,3,3,
2,2,1,1,
4
1
,3,,3,
4
1
,2,,2,
4
1
,1,,1,,
~
iiiiiiiiii
iiii
s
snewii
s
snewii
s
snewiinewii
yFyFyFyFyF
yFyFqcqcqcQc
+++++
++++= ∑∑∑=== (49)
To ensure that the correct fixed cost is taken into consideration in the cost function
and that at most one combination is chosen per bidder, the following constraints are
added to the GSP for each inactive bidder Ii ∈ ,
150
0
0
0
0
3,2,1 0
4,3,,4,2,,4,1,,123,
3,3,,3,2,,23,
2,3,,3,1,,13,
2,2,,2,1,,12,
1,,,,
≥−−−
≥−−
≥−−
≥−−
=≥−
newinewinewii
newinewii
newinewii
newinewii
knewiki
qqqMy
qqMy
qqMy
qqMy
kqMy
1123,23,13,12,3,2,1, ≤++++++ iiiiiii yyyyyyy
{ }1,0,,,,,, 123,23,13,12,3,2,1, ∈iiiiiii yyyyyyy
(50)
where M is a constant larger than any conceivable item quantity. In this example, M =
1000 was used. Note that these constraints leave open the possibility that yi,jkl assumes
the value of one, even though one or more of the items in the combination are zero.
For example, when items one and two assume a nonzero value for bidder i, either yi,12 =
1 or yi,123 = 1. However, since the fixed cost for a bundle including more items is always
larger than for a bundle with less items, the objective function will ensure that the yi,jkl,
which coincides exactly with the combination of nonzero item quantities, assumes the
value one.
The bundle of bids suggested by the GSM in the first round is presented in Table 24.
There are always multiple solutions, because the profit or loss in the bids (ei or si) can
be divided in an infinite number of ways between the bidders who are offered a bid
suggestion. We chose the solution where the estimated profit/loss is divided equally
among the bidders. Of course the true profits/losses of the bidders are not equal; we are
dividing the profit/loss calculated with the cost function estimates equally among the
bidders. Another approach would have been to divide the profit proportionately to the
size of the bid. The main effect that the choice of solution has in the auction, is that in
some cases it can cause the bidder to accept or reject a bid suggestion. E.g. if our cost
estimate is a bit too low, offering a bid suggestion in which the GSP thinks the bidder
will just break even, will not be good enough for the bidder and she will reject.
However, if some of the surplus in the auction is allocated to this bidder, it may
increase the bid price high enough so that she will accept the bid suggestion. Because
we assume we do not know the bidders’ true costs, it is difficult to know which cost
151
estimates are underestimated, and which way of dividing the surplus would be best.
Thus, we decided to start with dividing the (estimated) profit equally.
Table 24 Bids suggested by the GSM
Bid Item 1 Item 2 Item 3 Price (€)
x1,new 0 225 0 16,658
x5,new 75 225 0 21,426
x7,new 300 0 300 43,950
Bidders 1, 5 and 7 are suggested a bid, and their bids would team up with Bidder 6’s
initial bid (225; 150; 300; 49,356 €). The bidders all accept the new suggestions, so the
set of provisional winners is now Bidder 1, Bidder 5, Bidder 6 and Bidder 7, and the
total cost to the buyer is reduced to 131,390 €.
Next the cost functions are updated. The cost estimates for Bidders 3 and 4, who were
inactive but still not offered a new bid, are now lowered by 1%. The cost estimates of
the active bidders remain the same. The accepted bids offer no new information on the
bidders’ cost functions, because the bid prices are high enough to cover current
estimated costs. The reason for the decrease in Bidder 3’s and Bidder 4’s estimates is
that it is possible that we have overestimated their cost and that is the reason why they
were not offered a bid. Bidders 2, 8, 9 and 10 were active so they could not have
received bid suggestions regardless of their cost function estimates, and thus nothing is
done to the estimates of their cost functions.
Round 2
The GSP with the updated cost information is solved for the new set of inactive
bidders. The solution is presented in Table 25.
Table 25 Bids suggested by the GSM in Round 2
Bid Item 1 Item 2 Item 3 Price (€)
x4,new 300 225 225 53,426
x8,new 0 150 225 27,736
This time the GSP suggests bids to Bidder 4 and Bidder 8, which would team up with
the initial bids of Bidder 5 (x51 = 1) and Bidder 7 (x71 = 1). The GSP thinks the bids will
result in a loss of 474 € for both bidders. However, both bidders accept the suggested
bids. The GSP has overestimated their costs, so the true cost of producing the proposed
152
bundles is below the bid prices. Now we have new information on the costs of Bidders
4 and 8, and their cost estimates are updated. The cost estimates for inactive bidders
who did not receive a bid suggestion (Bidders 2, 3, 9 and 10) are lowered by 1%, and
the estimates for bidders who previously were active (Bidder 1, 5, 6 and 7), out of
which Bidders 5 and 7 are still active, remain untouched. The new cost function
parameters are depicted in Table 26.
Table 26 Cost function parameters after Round 2
Bidder 1 Bidder 2 Bidder 3 Bidder 4 Bidder 5 Bidder 6 Bidder 7 Bidder 8 Bidder 9 Bidder 10
F i 2222.2 2200.0 2178.0 2200.0 2222.2 2222.2 2222.2 2222.2 2200.0 2200.0
F ij 3333.3 3300.0 3267.0 3300.0 3333.3 3333.3 3333.3 3333.3 3300.0 3300.0
F 123 4444.4 4400.0 4356.0 4400.0 4444.4 4444.4 4444.4 4444.4 4400.0 4400.0
c 1 66.7 66.0 65.3 64.4 66.7 66.7 65.9 66.7 63.8 66.0
c 2 61.6 66.0 65.3 66.0 54.5 66.7 66.7 66.7 66.0 66.0
c 3 66.7 63.5 65.3 66.0 66.7 66.4 66.7 59.1 66.0 66.0
Round 3
First the GSP suggests a bid only for Bidder 10, and that bid (0; 150; 225; 25,160 €) is
not acceptable to her. Also according to the estimated cost the bid is not profitable, and
therefore we do not receive more information on B10’s cost function. The cost
estimates of all the other inactive Bidders (1, 2, 3, 6 and 9) are lowered by 1% and the
GSP is solved again. This time the GSP suggests bids for Bidder 1: (150; 225; 0; 26,139
€), Bidder 6: (225; 0; 300; 37,075 €) and Bidder 9: (225; 150; 150; 37,379 €), which
would team up with Bidder 5’s initial bid resulting in a total cost to the buyer of
126,187 €. All the three bidders find the suggestions profitable, even though the GSP
again thinks the bids would result in losses. The cost estimates are updated for all
bidders, except the ones who are active.
Round 4
The first solution of the GSP provides only a suggestion for Bidder 3 (225; 150; 150;
34,855 €), which is not accepted. Thereafter, the GSP is solved seven times, and each
time at least one bidder who is suggested a bid declines the suggestion, vetoing the
“group” bid. The accepted bids are added to the bid stream, but the total cost to the
buyer does not decline by the required 2% because the complementary bid(s) was not
accepted. After each GSP solution the cost functions are updated. Finally, the ninth
153
iteration produces bid suggestions to Bidder 4: (300; 0; 150; 30,515 €) and Bidder 7:
(300; 150; 300; 50,897 €), which are both profitable for the bidders.
After this round, the GSP does not find bid combinations that would be profitable for
all bidders. The auction ends. Going back to the bidders’ cost functions we can
conclude that the inactive bidders have such high costs that they could not afford to
decrease the total cost to the buyer by the required 2%. The active bidders could have
afforded to, but they did not have an incentive to do so, as they were already among the
provisional winners. The final bid stream and the winning bids are presented in Table
27.
Table 27 Bid stream and solution of the WDP after Round 4
Bid Item 1 Item 2 Item 3 Price (€) Status
x11 0 225 150 27,199 1
x21 150 0 150 22,954 0
x31 225 225 150 45,990 0
x41 0 225 0 17,613 0
x51 0 225 150 25,594 1
x61 225 150 300 49,356 0
x71 300 0 0 22,007 0
x81 0 225 225 33,210 0
x91 225 150 0 27,831 0
x10,1 225 225 225 50,076 0
x12 0 225 0 16,658 0
x52 75 225 0 21,426 0
x72 300 0 300 43,950 0
x42 300 225 225 53,426 0
x82 0 150 225 27,736 0
x13 150 225 0 26,139 0
x62 225 0 300 37,075 0
x92 225 150 150 37,379 0
x43 300 150 225 46,781 0
x44 300 225 225 50,969 0
x73 300 150 0 31,938 0
x45 300 0 0 19,666 1
x74 300 150 300 52,385 0
x46 300 150 0 30,561 0
x10,2 225 225 225 46,199 0
x47 300 0 0 19,778 0
x48 300 0 0 19,824 0
x49 300 0 150 30,515 0
x75 300 150 300 50,897 1
154
The total cost to the buyer is 123,356 €. In fact, the total cost decreased by more than
2% from Round 3. The previously inactive bid x45 made by Bidder 4 was more
advantageous to the buyer now that a good match entered (x75) the auction. The
estimated profit for Bidder 4 was larger from the newest bid (x49) and that is why the
GSP favored that one (it looks at the auction from the bidders’ perspective), but it is the
WDP that determines the set of winners. The mark-up in Bidder 4’s bid x45 is higher
than in bid x49. Hence, the bidder may be happier with this outcome.
10.3.3 Comparison with the Efficient Allocation
In order to see how well the GSM performed, we wanted to compare the final
allocation of the auction to the efficient auction. The efficient allocation can be solved,
because we know the bidders’ cost functions. It is presented in Table 28.
Table 28 The efficient allocation of the example auction
Bidder Item 1 Item 2 Item 3
B1
0 75 150
B4 300 0 0 B5 0 225 150
B7 300 300 300
Comparing the efficient allocation with the actual auction outcome (in Table 27) it is
evident that the auction came close to the efficient allocation. All the bidders are the
same, as are the items they bid on. The only differences are in two item quantities. If
we compare the production costs of the efficient allocation (113,666 €) and the
winning allocation (114,057 €), the difference (391 €) is very small.
10.3.4 Comparison with the QSM Auction
The auction presented in the example above was also run through using the QSM to
support the bidders instead of the GSM. All the auction parameters (demand, number
of bidders, bidders’ cost functions and initial bids) were kept the same, but instead of
using the GSM, the inactive bidders used the QSM, and placed bids based on the bid
suggestions made by the QSM. The final allocation of the QSM auction is presented
in Table 29.
155
Table 29 The final allocation of the auction with QSM
Bidder Item 1 Item 2 Item 3
B4
225 150 0
B5 0 150 150 B7 0 225 150
B8 150 0 150 B9 225 75 150
The combined production cost of the winning bidders is 121,385 €, which is 7,719 €,
or 6.8% higher than the efficient production cost, and 6.4% higher than the production
cost of the winning allocation of the GSM auction. There are some bidders among the
winners, who are not efficient and should not be there (Bidders 8 and 9), and one
efficient bidder (Bidder 1) is not among the winners. Also, the number of winning bids
in the final allocation exceeds that of the efficient allocation, which causes the bidders
to incur unnecessary fixed costs. The difference between the GSM and the QSM is
very clear, even though the efficient allocation consisted of only four bids. When the
number of bids increases, one can expect the difference in the efficiency of the QSM
auction and the GSM auction to increase.
In a way, the GSM can be considered as a generalization of the QSM. As a special
case, the GSM will support a single bidder when it finds the optimal course of action,
but it also provides the possibility of supporting any combination of bidders in each
iteration. Hence, it has a substantially higher flexibility to improve the allocations. The
benefits that can be obtained from this flexibility are, naturally, expected to increase as
the number of bids in the efficient allocation increases. Also, the GSM alleviates the
threshold problem and the extended puzzle problem, since it can offer bid suggestions
to a group of “local” or “glocal” bidders to help them outbid a “global” bidder. Thus,
the GSM should improve the allocative efficiency of the final allocation compared to
the QSM in cases in which the efficient allocation consists of three bids or more.
156
VI COMBIAUCTION – IMPLEMENTING THE QUANTITY
SUPPORT MECHANISM IN PRACTICE
The ultimate goal of our research project is to implement the bidder support tools in
practice. Thus, we have designed and developed an Internet-based auction system, the
CombiAuction, which enables combinatorial bidding, and to which the Quantity
Support Mechanism (presented in Chapter 7) and the Price Support tool of Teich et
al. (2001 and 2006) (formulation in section 8.2.3) have been integrated. Although per
se insufficient, the Price Support tool is a good complement for the QSM. Price
Support is faster to use, since it usually gives fewer suggestions than the QSM. Also, it
is a helpful tool in situations in which the bidder has placed bids for good item
combinations but the prices are too high.
The following chapter presents the CombiAuction system and the user interface. The
auctions system has been tested with human subjects. The purpose of the tests was to
study the feasibility of our auction system and the usability of the user interface, but
also to study how easily people grasp the somewhat complex idea of combinatorial
auction, and what kind of bidding strategies they use. Chapter 12 presents the
experiment design and the results of the experiment. The contribution of this part of
the thesis includes the implementation of the QSM and the design of the laboratory
experiment. The most important contributions, though, come from the observations
related bidder behavior. I present a classification of bidders according to their bidding
strategies, and discuss the cognitive challenges of combinatorial auctions. Especially,
analyzing the strategies used by the bidders in choosing the bids (prices and item
quantities) is interesting, since identification of such strategies has not been done
before.
157
11 THE COMBIAUCTION SYSTEM
The CombiAuction System is an Internet application designed by our research group20
and developed by Valtteri Ervasti. An earlier application, NegotiAuction, had been
developed by Teich et al. (2001, 2006) and used to run single-item, multi-unit
auctions. In the development of CombiAuction, the database and the underlying data
model were taken from NegotiAuction and developed further to suit the needs of
combinatorial auctions. Aside from this, improvements in technology since
NegotiAuction, and the fundamental differences between combinatorial and single-
item auctions made it necessary make major changes in the application.
Combiauction is designed as a web application and written in Java according to the
MVC (Model-View-Controller) architectural pattern. At the center of the application,
a calculation package is written to handle all the calculations of the WDP and the
QSP. A commercial software package (LINDO) is used to solve the problems. The data
is stored in a MySQL database, and for access to the database, Hibernate is used in the
persistence layer. For the Control and View parts of the architectural pattern,
CombiAuction implements the widely used Apache Struts framework. The user
interface consists of pages written in JSP, with the Struts tag libraries heavily utilized.
The finished application is deployed in Apache Tomcat, a web container. Finally, the
web container and the database were installed on a virtual server. With this
configuration, CombiAuction can be accessed through the Internet using any web
browser.
11.1 Organizing an Auction in the CombiAuction System
All auctions run on the CombiAuction system are continuous auctions, that is, the
WDP is solved after each incoming bid. Some of the auction design parameters and
auction rules are predetermined, but some the auction owner can specify.
CombiAuction supports the organization of multiple-unit combinatorial auctions in
both forward and reverse formats. Thus far, the only attributes the bidders can bid on
20 Main contributors in the design apart from myself were professors Hannele Wallenius and Jyrki Wallenius and Valtteri Ervasti.
158
are the item quantities and bid price. Multiple attributes (quality, delivery terms, etc.)
could be incorporated into the auction through “pricing out” as in NegotiAuction
(Teich et al., 2001 and 2006). However, “pricing out” has not been implemented yet,
since the focus has been on the combinatorial aspects of the auctions. In the following
I will present the kinds of auctions one can organize in the CombiAuction system.
A new auction can be created by clicking on the link “Create a new Auction” in the
menu bar. A new auction is created in four steps. First the owner provides the auction a
name and a brief description, chooses the direction of the auction (reverse or forward),
and specifies the number of items to be sold/bought (see Figure 3). Assume that in this
example, the owner wants to create a reverse auction.
Figure 3 Screen view from Step 1 of the auction creation process
In the second step (Figure 4) the auction owner specifies the items, and the number
and type (kilograms, liters, tones, units, etc.) of units of each item to be bought. The
owner can give a description for each item, if she wants to, and specify a reservation
price for the whole bundle. The reservation price is not revealed to the bidders. The
owner can also specify a maximum number of units of each item that is allowed in
each bid. If the owner restricts the maximum quantity offered by each bidder to be less
than the total demand, the owner can make sure that at least two bidders are chosen as
159
suppliers for that item. This way the buyer can decrease, for instance, her dependence
on any particular supplier.
Figure 4 Screen view from Step 2 of the auction creation process
In the third step (Figure 5) the auction owner sets the opening and closing times for
the auction. Although the owner specifies a closing time, the auction closing rule is
flexible. Every time someone places a bid within the last 10 minutes of the auction, the
closing time is pushed back by 10 minutes. The closing time is pushed back until
bidding stops. This “soft” closing rule is often used to prevent bidders from sniping.
Besides the auction opening and closing times, the auction owner specifies the
minimum bid decrement (increment in a forward auction) by which the total cost to
the buyer must decrease when the provisionally winning allocation changes. As all the
auctions are continuous, a round is defined as a change in the provisionally winning
allocation. The auction owner can choose whether the bid support tools are available
or not. More specifically, choosing “full quantity support” means that both quantity
160
and price support are available. The two other alternatives are to enable only price
support, or to offer no support at all. The owner can either organize an open-entry
auction to which everyone registered in the CombiAuction system can participate in,
or restrict participation by requiring that the owner has to confirm the bidders’ request
to participate (“Entry by confirmation”). The last design element to choose is to specify
what information is revealed to the bidders (bid visibility). “Closed bidding” here refers
to what I have called a semi-sealed-bid auction in this thesis. That means that bidders
see their own bids and their statuses (active or inactive), but not the number or content
of other bids. “Open bidding” means that all bids placed in the auction and their
statuses are visible to all participants. If choosing an open auction, the owner can select
whether the bidders’ identities are kept hidden (second alternative), or revealed to
everybody (third alternative). Regardless of the type of bid visibility chosen, the auction
owner will always see all the bids and know which bidder placed them.
Figure 5 Screen view from Step 3 of the auction creation process
161
The last step simply pools together all the design information specified by the owner
for a final confirmation before the auction is created (Figure 6).
Figure 6 Screen view from Step 4 of the auction creation process
After the auction is created, the auction owner can monitor the progress of the auction
from the owner’s auction home page (Figure 7). The owner’s auction home page
contains four boxes, each containing information related to the auction.
162
Figure 7 Screen view of the auction owner's auction home page
The “Auction Information” box reminds the auction owner of the opening and closing
times, and the new closing time will also be updated there in case the closing time is
pushed back due to late bidding. The owner can also check the current total cost from
the “Auction Information” box. The link “Auction parameters” lets the owner review
some of the auction parameters specified in Steps 2 and 3 of the auction creation
process. Through the link the owner can make changes in the bid decrement, the
reservation price, the auction closing time, and the minimum and maximum item
quantities allowed in each bid. The link “Auction bidders” returns a list of bidders
whose request to participate the owner has accepted. The owner can also conveniently
send a message to all participants through the messaging system built in the
CombiAuction system by clicking on “Send message to all bidders.”
163
The “Auction Items” box lists the items in the auction and item quantities demanded
by the owner (buyer). By clicking on the item names the owner can check the item
description given originally in Step 2 of the auction creation process. The box also
informs the owner of the sum of item quantities in the currently active bids. As can be
seen from screen shot in Figure 7, currently the demand is met, and there is no excess
supply. The auction system does not allow shortages of item supply, but it is acceptable
to have excess supply of the items in the active bids as long as the total cost is the lowest
possible and the reserve price is not exceeded. The “Auction Items” box also reports the
current shadow prices the QSM uses to solve for the bid suggestions. In case the
shadow price for some item is zero, the shadow price is set to one because otherwise
the item would have a zero cost and disappear from the objective function of the QSP.
The “Incoming bid requests” box lists all the bidders wishing to participate in the
auction. The owner can either confirm or reject the bidder’s request. The “Incoming
bid requests” box disappears when there are no requests.
The “Auction Bids” box lists all the bids placed in the auction with the newest bid on
top. The owner can see the time the bid was placed, the bidders’ names, the content of
the bids, and the bid statuses. The statuses (active or inactive) will change to “won” or
“lost” once the auction closes. The owner can “lock” some of the bids, which means
that those bids are guaranteed winners regardless of what other bids may enter the
auction. This feature was incorporated already in NegotiAuction, and has been
included in the CombiAuction as well. The ability to lock some bids prior to the end of
the auction brings the auction closer to a negotiation. For instance, a supplier could be
pressed for time and promise to make a good offer under the condition that the buyer
makes an immediate decision whether to accept it or not. The owner can also “disable”
bids, which prevents the bids from becoming active, or “delete” them entirely from the
bid stream. By clicking on the names of the bidders the auction owner can send
messages to the bidders.
11.2 Bidding in the CombiAuction System
A bidder registered in the CombiAuction system can browse through the list of open
auctions and auctions scheduled to open in the future (Figure 8).
164
Figure 8 Screen view of the list of open and scheduled auctions.
By clicking on the name of the auction, the bidder can look at the details of the
auction (Figure 9).
Figure 9 Screen view of the auction information page
The bidder can see the items and item quantities demanded by the owner (buyer), and
by clicking on the names of the items, the bidder can see the item descriptions given
by the owner. If the bidder wishes to participate in the auction, she will click on the
“Participate in the auction”. If it is an open-entry auction, the auction will be directly
added to the bidder’s “My Auctions” page (Figure 10). If participation requires the
auction owner’s participation, clicking on the link will send a request to the owner.
After the owner confirms the request, the auction is added to the bidder’s “My
Auctions” page. Bidders can enter open auctions at any point in time; they are not
165
required to be present from the beginning to the end. Also, because of this free entry,
no activity rules are enforced in the CombiAuction.
Figure 10 Screen view of the bidder's "My Auctions" site
Once a participant in an auction, the bidder is directed to the bidder’s auction home
page (Figure 11), where she can review the details of the auction (which are the same
presented in on the auction information page for prospective participants in Figure 9),
monitor her status in the auction, and submit new bids.
166
Figure 11 Screen view of the bidder's auction home page
The bidder can submit bids in three ways. First, the bidder can place a new, “self-
made” bid by filling out the item quantities and the bid price in the dialog box in the
lower left hand corner. In CombiAuction, all bids are allowed to enter the bid stream.
In other words, the bidder is not required to place a bid that becomes active upon
submission. Also inactive bids are accepted. This is to help overcome the threshold
problem. If the bidder only wants to change the price on one of her existing bids, she
can use the “Reprice” option next to the bid. Thirdly, the bidders can use the two
support tools, the Quantity Support Mechanism and the Price Support tool to help
them submit bids – if the auction owner has included the tools in the auction. Notice,
however, that the bidder support tools are never available at the beginning of the
167
auction. There has to be some bids in the bid stream before the support tools can find
bids that can become active.
When the support tools are available, the link appears in the “Auction Information”
box. The link leads to the support dialog page (Figure 12).
Figure 12 Screen view from the support tool dialog page
The link to the price support tool is at the bottom of the page. Simply by clicking on
the link “Find new prices”, the auction system will offer the bidder a list of those of her
inactive bids that can become active with a lower price (Figure 13).
168
Figure 13 Screen view of the Price Support tool
In this example, only one of the bidder’s existing bids can become active. The bidder
has the option of submitting this bid suggestion, or returning to the auction home page
without placing a bid. If the bidder submits the bid, it is added to the bid stream and it
will appear under “Your Bids” on the bidder’s auction home page.
The Quantity Support tool (upper box in Figure 12) allows the bidder to specify some
constraints on the suggestions the QSM will suggest. The bidder can restrict the
maximum amount of each item, which is convenient if the bidder is operating under a
capacity constraint. The bidder can also specify a lower limit for item quantities and
the bid price in case she is interested only in bids of a certain size, or wants to bid on
some items in particular. If the bidder does not insert any specifications, the QSM will
use zero as the default value for the lower limits and the bid price, and the maximum
allowed item quantity defined by the auction owner as the default value for the upper
limit. The shortlist in the QSM is the “full” short list described in section 8.2.2.
The bid suggestions offered by the QSM are offered to the bidder (Figure 14), and the
bidder can evaluate the suggestions and submit the one that is the most profitable to
her (if any). If the bidder submits a bid, it is added to the bid stream and it will appear
under “Your Bids” on the bidder’s auction home page.
169
Figure 14 Screen view of the bid suggestion made by the QSM
The auction continues until the closing time, or until no new bids have been
submitted for 10 minutes.
170
12 TESTING THE COMBIAUCTION SYSTEM21
The CombiAuction system was tested in an experiment with human subjects. The
objectives of the experiment were to 1) test the feasibility and usability of the
CombiAuction system, 2) study how the auction outcomes (efficiency and total cost to
buyer) compare with those obtained in the second simulation study, 3) study what kind
of strategies bidders use, and 4) how human users understand the concept of a
combinatorial auction and the support tools.
12.1 Experiment Set-Up
In the experiment, two different auction settings were used. The designs for both
settings were chosen from the second simulation study (presented in section 8.2) to
allow for direct comparison to the simulated auctions. Also all the parameter values
were set to be the same as in the simulations studies (5 items, 600 units of each item,
2% decrement). In the simulation study we tested 48 different designs, which could not
have been reproduced with a limited number of human subjects. Thus, I chose two of
the designs, one design with equal capacities, and one with unequal capacities, since
based on the simulation study it was the bidders’ capacities which had the biggest
impact on the efficiency of the auction outcome.
For the first experiment setting – auction A – I chose an equal capacities auction with
15 bidders, normal economies of scope, and full shortlist. The initial bids were created
based on advantage in variable costs (Bid1). In the simulation study, the efficient
allocation changed from replication to another (there were 50 replications of each
design). For the laboratory experiment I chose only one replication, that is, one set of
cost functions. Using the same cost functions for each group of participants allowed for
better comparisons between the groups. I deliberately chose a replication which had
ended in the efficient allocation in the simulation in order to see if the auctions would
end in the same, efficient allocation with real users as well.
21 I am solely responsible for the design and conduct of the laboratory experiment as well as for the analysis of the results.
171
For the second setting (auction B) I chose the corresponding unequal capacities
auction (15 bidders, normal economies of scope, full shortlist and initial bids created
based on Bid1). However, this time I chose a replication in which the winning
allocation was inefficient (winners’ combined production costs were 9.2% above the
costs of the efficient allocation), in order to study whether human users could direct
the auction to a more efficient allocation or not.
Experiment Participants
In total 74 students – both undergraduate and graduate students – participated in the
experiment. The students were all participants of a course on managerial economics at
Helsinki University of Technology. Thus, the students were already familiar with the
concepts of production and cost functions, and economies of scale and scope.
However, only 14 of them had participated in online auctions before, and the
combinatorial auction was an unfamiliar concept to all of them. Each student was
required to participate in two auctions: first in one A auction and then in one B
auction. All 74 participants bid in the first (A) auction, but only 69 of them bid in the
second (B) auction. The students were rewarded by giving them points that counted
towards their final grade from the course. In the beginning each student received 5
points, and they were rewarded extra points for playing well (winning with a profit
larger than their counterparts), and for answering a post-experiment questionnaire. In
case the students did not perform as expected (for example, placed unprofitable bids or
failed to win when they had the chance to), points were deducted from them. The
maximum score attainable was 11 points, which is 11% of the final grade from the
course. Giving credit for participation is an easy and cheap way of motivating students
to participate in laboratory experiments.22
Prior to the experiment, all participants were required to participate in a briefing
session. The briefing session contained some general theory on combinatorial auctions.
Also, the students were briefed on how to use the CombiAuction system, how the
support tools work, and how to use the cost function parameters given to them on an
22 Also Bichler (2000) gave students credit for participating in his experiments.
172
Excel sheet to calculate their costs for each bundle. It was explained to the students
how the winners are determined in a reverse combinatorial auction, and what
principles the QSM is based on (and why it is not available at the beginning of the
auction). However, no mathematical notation or formulations of the WDP or the QSP
were presented.
The Organization of the Experiment
The experiment was organized in four sessions during four consecutive days so that
there were two sessions (days) for A auctions: A1 and A2, and two sessions (days) for B
auctions: B1 and B2. The reason for having two sessions for both auctions was to offer
the students the possibility to choose the days that suited their schedule the best. Each
auction lasted for 23,5h (or longer if the closing time was extended). The participants
were physically in different locations during the auction, but participated over the
Internet. The duration of 23,5h is longer than the 1-4h duration usually used in
laboratory experiments (in which the participants usually are in the same place at the
same time). However, I chose such a long duration to better simulate an actual online
auction in which the endogenous arrival of bidders is a key characteristic. Real online
auctions usually last for several days or weeks, partly in order to give time for potential
buyers to find the auction. However, since in this experiment the bidder pool was
predefined and the prospective bidders were all knowledgeable about the auction,
there was no need to extend the auction longer than was needed to be able to observe
different bidding strategies. The auctions began at 5pm and the scheduled closing time
was 4.30 pm the following day. This way the bidders had essentially two days time to
place bids, which I anticipated to be enough to separate early bidders from late bidders.
Even though the simulated auctions, from which I borrowed the design parameters,
cost functions and capacities, all had 15 bidders, I chose to divide the students into
smaller groups. There were two reasons for this. Firstly, that way I could get more
replications of the auctions. Then, if for some reason some of the auction outcomes
were distorted (e.g. due to mistakes made by the bidders) there would be enough data
left to analyze. Secondly, I thought it would be more rewarding for the students if a
bigger portion of them could win at least once during the experiment. For the A
173
auctions, 46 students enrolled in session A1, and 28 students in A2. Thus, I divided the
A1 participants into 9 groups of 5 students, and one group of 6 students, and A2
participants into 4 groups of 7 students. Five students were enough to guarantee
sufficient competition in equal capacities auctions in which the efficient allocation
consisted of two bidders. Even if bidders placed bids for smaller bundles than
maximum capacity, it would be very unlikely that all five bidders would be provisional
winners simultaneously. For the unequal capacities auctions five bidders might not
have been enough, since the bidders’ capacities were smaller than in the equal
capacities auctions, and more bidders would always be needed in the winning
allocations. Thus, the 40 participants in the B1 session were divided into 4 groups of 7
students and 2 groups of 6 students, and the 24 participants in the B2 session were
divided into 4 groups of 7 students and one group of 6 students.
The students were assigned cost functions to identify them. From the 15 bidders in the
simulations I chose 5-7 bidders to be assigned as identities to the students. The same
bidder identities (cost functions) were used in all the groups to allow for direct
comparisons between groups. This way I could compare the auction outcomes
(efficiency and total cost to the buyer) to see if they differed from one group to another
– even though the starting points were equal. Also, when the same bidder identities
were used in each group, the bidders’ performance could be compared with
corresponding bidders in other groups. This comparison determined the participants
reward points. For the equal capacities (A) auctions with five bidders I chose the cost
functions of the two bidders, who formed the efficient allocation (Bidders 12 and 15) 23.
In addition I chose the cost function of one bidder (Bidder 10) whose cost for the
efficient allocation bid (300, 300, 300, 300, 300) was close enough to the efficient
bidders so that a “pseudo efficient” allocation could be reached. A pseudo efficient
allocation was defined as an allocation which was not efficient, but in which the total
cost to the buyer was within the 2% decrement of the efficient production costs. In the
pseudo efficient allocation the efficient bidder(s) cannot afford to submit new bids
because in order to reduce the buyer’s cost by 2% they would have to incur a loss. The
23 The bidders’ numbers refer to their number in the simulated auction, which makes is easy to keep track of the bidders’ costs.
174
two remaining bidders whom I chose (Bidders 5 and 8) had higher costs, and they
should not be among the winners. When there were 6 or 7 participants in the auction,
the extra bidders were assigned costs that could also result in pseudo efficient outcomes
(Bidders 1 and 11), but who had higher costs than Bidders 12, 15 and 10. The bidders’
cost function parameters are presented in Appendix 4.
In the unequal capacities (B) auctions I included the bidders from the efficient
allocation (Bidders 3, 8 and 13), the bidders who won in the simulated auction
(Bidders 4 and 15 in addition to Bidder 3), and the remaining one or two bidders were
chosen at random to be Bidders 1 and 9. In the unequal capacities case it would have
been impossible to try to deduce which bidders could create a pseudo efficient
allocation, since the item combinations in the winning bids also would have to change
due to the different capacities (in the equal capacities case the winning bids would
almost always be for half the total demand). The bidders’ cost function parameters and
capacities are presented in Appendix 4.
The organization of the experiment is summarized in Figure 15. The participants were
divided into A1, A2, B1 and B1 sessions according to their preferences. I then further
divided them into smaller auction groups within each session. In each auction there
were the same set of cost functions (bidder identities) given to the participants. The
cost functions were assigned randomly in the A auctions, but in the B auctions I tried
to give better cost function to those who had received the worst ones in the A auction.
This way I tried to give everyone equal chances to obtain extra points. Everybody
participated first in one A auction and then in one B auction. The equal capacities (A)
auctions were slightly simpler bidding environments, hence they also served as a
practice session for the B auctions. The participants were grouped differently in the A
and B auctions. The idea was to have bidders bid against new competitors – who
maybe used a different strategy – in the second auction. Also, due to the participants’
diverse preferences and different group sizes, it would have been impossible to
maintain the same groups in both A and B auctions.
175
AUCTION EXPERIMENT
A Auctions B Auctions
A1 (Mon-Tue)
A1-a
Bidder 5Bidder 8Bidder 10Bidder 12Bidder 15
A1-hBidder 8Bidder 10Bidder 12Bidder 15
Bidder 5
A1-iBidder 8Bidder 10Bidder 12Bidder 15
Bidder 5Bidder 1
A2 (Tue-Wed)
A2-a
Bidder 5Bidder 8Bidder 10Bidder 11Bidder 12
Bidder 1
Bidder 15
A2-d
Bidder 5Bidder 8Bidder 10Bidder 11Bidder 12
Bidder 1
Bidder 15
B2 (Thu-Fri)
B2-a
Bidder 3Bidder 4Bidder 8Bidder 9Bidder 13
Bidder 1
Bidder 15
B2-d
Bidder 3Bidder 4Bidder 8Bidder 9Bidder 13
Bidder 1
Bidder 15
B2-eBidder 4Bidder 8Bidder 13Bidder 15
Bidder 3Bidder 1
B1 (Wed-Thu)
B1-a
Bidder 3Bidder 4Bidder 8Bidder 9Bidder 13
Bidder 1
Bidder 15
B1-d
Bidder 3Bidder 4Bidder 8Bidder 9Bidder 13
Bidder 1
Bidder 15
B1-eBidder 4Bidder 8Bidder 13Bidder 15
Bidder 3Bidder 1
B1-fBidder 4Bidder 8Bidder 13Bidder 15
Bidder 3Bidder 1
Figure 15 The organization of the laboratory experiment
12.2 Results of the Auction Experiment and Observations
In this chapter I will first compare the outcomes of the laboratory auctions to the
simulated auctions. Thereafter I will discuss bidder behavior (strategies), bidders’
understanding of the combinatorial auction and the support tools, and the usability of
the auction system. The discussion is based on my observations from the auctions and
the answers the participants gave to a post-experiment questionnaire after completing
the auctions. The post-experiment questionnaire can be found in Appendix 5. Out of
the 74 participants, 66 returned the post-experiment questionnaire.
12.2.1 Efficiency of the Final Allocation and Total Cost to Buyer
The efficiency of the final allocation and the total cost to the buyer were the key
performance measures studied in the second simulation study. Thus, they served as a
convenient starting point for the analysis of the laboratory experiment as well.
176
Equal Capacities (A) Auctions
The simulated version of the equal capacities (A) auction ended in an efficient
allocation. Table 30 presents the winning allocations of the simulation, and of the
laboratory experiments. Auctions A1-d, A2-a, and A2-c were left out from the table,
because in those auctions one of the winning bidders had mistakenly placed a bid with
negative profit, which distorted the final outcome. Thus, studying the efficiency or the
total cost to the buyer in these auctions does not make sense. The efficiency could be
really bad and the total cost to the buyer still low because of the unprofitable bids.
These auctions are taken into consideration in the next sections, which discuss bidder
behavior.
Table 30 Winning bidders and bids in the equal capacities (A) auctions
Bidder B 12 B 15 B 8 B 15 B 10 B 15 B 12 B 15 B 12 B 15 B 5 B 12Item 1 300 300 300 300 300 300 300 300 300 300 300 300Item 2 300 300 300 300 300 300 300 300 300 300 300 300
Item 3 300 300 300 300 300 300 300 300 300 300 300 300Item 4 300 300 300 300 300 300 300 300 300 300 300 300Item 5 300 300 300 300 300 300 300 300 300 300 300 300
Price 142,214 138,000 145,100 137,349 141,979 140,138 158,232 140,000 150,000 139,000 147,000 140,000
Prod. cost 139,417 137,348 145,092 137,348 141,979 137,348 139,417 137,348 139,417 137,348 146,881 139,417Profit 2,797 652 8 1 0 2,790 18,815 2,652 10,583 1,652 119 583
A1-f A1-gA1-c A1-eA1-a A1-b
Bidder B 5 B 15 B 12 B 15 B 11 B 12 B 15 B 10 B 15 B 12 B 15Item 1 300 300 300 300 300 300 0 300 300 300 300
Item 2 300 300 300 300 0 300 300 300 300 300 300Item 3 300 300 300 300 300 0 300 300 300 300 300
Item 4 300 300 300 300 300 300 0 300 300 300 300Item 5 300 300 300 300 300 300 0 300 300 300 300Price 147,924 144,042 143,000 137,493 118,141 110,643 59,000 141,979 138,344 142,982 142,152
Prod. cost 146,881 137,348 139,417 137,348 117,003 109,521 58,071 141,979 137,348 139,417 137,348Profit 1,043 6,693 3,583 145 1,138 1,122 929 0 996 3,565 4,804
A2-b A2-d SimulatedA1-h A1-i
Out of the 10 auctions presented in Table 30, four auctions ended in the efficient
allocation (A1-a, A1-e, A1-f and A1-i). In addition, two auctions ended in a “pseudo
efficient” allocation (A1-c and A2-d). In these two auctions the inactive efficient bidder
(Bidder 12) should have placed an unprofitable bid in order to decrease the total cost
to the buyer by the required 2%. Had the decrement been smaller, the true efficient
allocation might have been reached.
Out of the remaining four auctions, which ended up in an inefficient allocation, three
auctions (A1-b, A1-g and A1-h) were such that at least the efficient bidder would have
177
been able to become a winner with a profitable bid. It is possible that the reason for
them not placing the bid is that the QSM did not suggest it to them. Another possible
reason is that the “shortlist” provided by the QSM was too long, and the bidders did not
have time to calculate the profits for all the bid suggestions – or did not want to go
through the trouble. Some participants complained that towards the end of the
auction, when most bidders were logged into the system, the QSM became slow, and
that it took several minutes for it to provide a list of suggested bids. This most likely has
affected the auction outcomes. This could also explain why the efficient bidders failed
to place their winning bids. The auction A2-b on the other hand, seems to have
suffered from the threshold problem. Bidder 15 placed her bid early in the auction,
and directly with a relatively low profit. There bids belonging to the efficient allocation
were also in the bid stream, but because the price was very high, the incoming bidders
could not afford to team up with them, but rather teamed up with Bidder 15’s cheap
bid.
Besides efficiency, also the total cost to the buyer is of interest in the auction outcome
– especially for the buyer. Table 31 summarizes the total cost to the buyer (as the ratio
of total cost to buyer and efficient production cost), as well as the efficiency indicator
(see definition in section 8.2.4.2) and the efficiency status of the auctions.
Table 31 Total cost to buyer and efficiency indicators from the A auctions
A1-a A1-b A1-c A1-e A1-f A1-g A1-h A1-i A2-b A2-d Simul.
Total cost to buyer (ratio
to efficient cost)1.012 1.021 1.019 1.078 1.044 1.037 1.055 1.013 1.040 1.013 1.030
Efficiency indicator 1 1.021 1.009 1 1 1.034 1.027 1 1.028 1.009 1
Efficiency status eff. ineff.pseudo
eff.eff. eff. ineff. ineff. eff. ineff.
pseudo eff.
eff.
According to Table 31 two of the auctions (A1-e and A1-f), which ended in the efficient
allocation, resulted in a cost to the buyer higher than in the simulated auction. The
explanation to this is that actually in these two auctions (contrary to the simulated
auction), there were still some bidders who could have afforded to submit a
provisionally winning bid, but for some reason they did not. Their answers to the
questionnaire indicate that one problem was the slow performance of the QSM at the
end of the auction. One of the bidders had not even tried using the QSM, but resorted
178
to placing self-made bids, which did not become active. Interestingly, the two
remaining efficient auctions (in which other bidders could not have afforded to place a
winning bid), both pseudo efficient auctions and even one inefficient auction ended in
a total cost to the buyer which was lower than in the simulated auction. This is because
in the experiment, some bidders bid unnecessarily low. They placed bids for very low
profit margins already quite early in the auction, when it would not have been
necessary to become a provisional winner. Either the bidders were very risk averse and
tried to maximize their probability of winning, or they did not want to constantly
monitor the auction and keep bidding, or then they did not want to use the QSM (or it
was not available yet). The bidders’ behavior is discussed further in the following
sections (12.2.2 and 12.2.3).
All in all, the QSM seemed to have performed quite well in the hands of human users
in the equal capacities auctions. The efficient and pseudo efficient allocation was
reached 6 times out of 10, and the efficiency indicators of the inefficient final
allocations are quite small (Table 31). The heavy bidding activity right before the
auction closing caused some problems for some of the bidders. This, however, is a
computational or a server-related issue, and not due to the structure of the QSM. This
observation does indicate though that the shortlist should be kept relatively short. Even
with five bidders there can be dozens of bids in the bid stream, and hence the shortlist
can become quite long. The threshold problem in auction A2-b shows that the content
of bids in bid stream and the order in which bids are submitted affects the auction
outcome.
One source of discrepancy in the results of the experiment auctions and the
simulations is the fact that the real auctions differed from the simulations on a few
design issues. First of all, in the simulated auctions the bidders did not place self-made
bids after the QSM became available. This affects the content of the bid stream a lot –
especially since the bidders in the laboratory experiment were more eager to submit
self-made bids throughout the auction than we had anticipated. Secondly, in the
simulations the bidders always submitted the most profitable bid on the shortlist. This
is not necessarily always the case in the laboratory experiment, especially when the
shortlist is long. Unfortunately, the shortlists are not recorded in the auction system,
179
and hence there is no way to verify afterwards whether the bidders chose the profit
maximizing bid or not. The system will be changed for the next experiment. Thirdly,
all the bidders were present at the same time in the simulated auctions, hence all the
inactive bidders were as likely to be the one submitting the next bid. In the laboratory
experiment this was not the case, as some of the bidders waited as long as until the last
hour to start bidding. However, one should also keep in mind that there is a random
element in the simulated auction as well (the inefficient bidder using the QSM is
chosen randomly), but the auctions were not repeated with the same cost functions.
Had the same auction been simulated several times, the outcomes might not have
been the same every time.
Unequal Capacities (B) Auctions
For the unequal capacities (B) auctions I had chosen an auction which in the
simulation study had ended up in an inefficient allocation. Table 32 presents the
winning allocation of the simulation and of the laboratory experiments, and also the
efficient allocation. Auctions B1-e, B1-f, B2-a, B2-d. and B2-e were left out from the
table, because in those auctions one of the winning bidders had mistakenly placed a
bid with negative profit, which distorted the final outcome of these auctions. Also
auction B1-a was left out because one participant, who had been given the identity of
Bidder 3 (one of the efficient bidders) did not bid in the auction, even though she had
requested for participation, and had been accepted as a participant. These auctions are
considered in the sections on bidder behavior (sections 12.2.2 and 12.2.3).
Table 32 Winning bidders and bids in the unequal capacities (B) auctions
Bidder B 4 B 8 B 13 B 1 B 4 B 8 B 3 B 8 B 15 B 1 B 4 B 15Item 1 300 300 0 0 300 300 150 300 150 0 300 300Item 2 150 150 300 225 150 225 0 300 300 200 150 250Item 3 300 0 300 225 300 75 300 150 150 200 300 100Item 4 300 0 300 0 300 300 0 300 300 0 300 300Item 5 300 0 300 300 300 0 150 300 150 300 300 0Price 135,000 56,454 120,000 80,000 134,000 100,047 73,000 163,467 123,892 73,000 133,300 109,554Prod. cost 133,270 53,429 116,897 73,899 133,270 97,340 71,619 137,291 115,615 71,179 133,270 101,048Profit 1730 3,025 3,103 6,101 730 2,707 1,381 26,176 8,277 1,821 30 8,506
B1-a B1-b B1-c B1-d
180
Bidder B 3 B 8 B 13 B 3 B 8 B 13 B 3 B 4 B 15 B 3 B 8 B 13Item 1 300 300 0 300 300 0 300 0 300 300 300 0Item 2 300 0 300 300 0 300 300 0 300 300 0 300Item 3 300 0 300 150 150 300 300 150 150 300 0 300Item 4 0 300 300 0 300 300 150 150 300 0 300 300Item 5 0 300 300 0 300 300 150 300 150 0 300 300Price 91,437 87,266 130,138 79,067 106,000 130,000 133,401 69,491 130,067Prod. cost 85,432 86,943 116,897 76,980 100,546 116,897 124,101 67,690 124,077 85,432 86,943 116,897Profit 6,005 322 13,241 2,087 5,454 13,103 9,300 1,801 5,990
B2-b B2-c Simulation Efficient allocation
Interestingly, in two auctions (B2-b and B2-c) the winning bidders were the same as in
the efficient allocation; and in B2-b even the bids were identical to the efficient
allocation bids. According to the bidders’ answers in the questionnaire, it was the
support tools that helped guide the auction to the efficient allocation. One of the
winning bidders in the B2-b auction had used price support, and the other two quantity
support to find the winning bids. Looking at the other auctions it is easy to see that the
winning bids vary a lot from one auction to another. This, and the fact that in the
simulated auction the QSM did not find the efficient allocation, demonstrates the
significance of the bids the bidders place without the help of the QSM in shaping the
progress and ultimately the outcome of the auction. Also the fact that two of the three
winning bids in auction B2-c were placed before the QSM was available attests to the
significance of the initial bids in the bid stream.
A better comparison between the experiment auctions and the simulation can be done
by studying the total cost to the buyer and the efficiency of the final allocations (see
Table 33).
Table 33 Total cost to buyer and efficiency indicators from the B auctions
B1-a B1-b B1-c B1-d B2-b B2-c Simul.
Total cost to buyer
(ratio to efficient cost)1.077 1.086 1.246 1.092 1.067 1.089 1.151
Efficiency indicator 1.050 1.053 1.122 1.056 1 1.018 1.092
Both the total cost to the buyer and the efficiency of the winning allocation are better
in the laboratory experiments than in the simulated auction with the exception of
auction B1-c. The inefficient winning allocation and high cost to buyer in auction B1-c
can be explained through lack of competition. There were seven bidders registered in
the auction, but only five of them placed bids. In all the other auctions in Table 33 all
181
bidders registered for the auctions also placed bids. In addition, out of those five
bidders who submitted bids in the B1-c auction, one bidder (Bidder 13) bid only at the
beginning of the auction. She placed several bids before anyone else had entered the
auction, but did not bid at all after other bidders started bidding. Thus, as the closing
approached, there were three active bidders and only one inactive bidder attempting to
submit bids.
An explanation for the better outcomes of the laboratory auctions compared to the
simulated auction can be found by observing the bidders’ bidding behavior in the
auctions. In auctions B1-a, B1-b, and B1-d two out of the three winning bids were
placed without the help of the support tools. Also, these bids were placed relatively
early in the auction (sometimes already before the demand was fulfilled), and the
profits in these bids were relatively low. In other words, the bidders were selling
themselves short not knowing that much higher profits could be gained. In B1-d, the
highest profit was made by the bidder, who was the last to bid – and who was the only
one of the winning bidders to have used the QSM. In B2-c the situation was the same
in the sense that again two of the three winning bids were placed among the first bids
in the auction. However, in this auction the profit margins in these two bids were high.
Luckily for these two bidders, they were both efficient bidders and their bids were close
to the efficient allocation bids – and the third efficient bidder who entered the auction
later was offered a good complementary bid by the QSM. Thus, even with high profits
in the winning bids, the total cost to the buyer remained at a reasonable level, and
much lower than in the simulated auction.
Based on the above-mentioned observations, it would seem that one explanation for the
better outcomes of the B auctions compared to the simulated auction is the fact that
bidders placed a lot more self-made bids than in the simulated auction. This widened
the possibilities for the QSM to find profitable bid suggestions. It would appear,
though, that a more powerful explanation is bidders’ unnecessarily low profits in their
initial bids. It is not easy to try to figure out a good strategy in combinatorial auctions –
especially in auctions where the efficient allocation is not necessarily reached, since
there is an element of luck involved. Figuring out a strategy is even more difficult for
inexperienced bidders, who are not familiar with the characteristics of combinatorial
182
auctions. In this experiment, the experience from the A auctions, in which profits were
lower in the winning bids, can have guided the bidders’ behavior in the B auctions.
12.2.2 Bidding Strategies
Bidding strategies in online auctions have been studied by Bapna et al. (2000, 2003
and 2004), Shah et al. (2003) and Puro (2009). They identify attributes that can be
used to categorize bidders’ behavior into distinct strategies. I will first review the
strategies they have identified. These strategies are not directly applicable in the
combinatorial auctions held in CombiAuction, and hence I will discuss the
peculiarities of the CombiAuction auctions before analyzing the bidders’ strategies in
the laboratory experiments.
12.2.2.1 Strategies Identified in Literature
Bapna et al. (2004) identify four bidder types each following their own strategy:
evaluators, opportunists, participants, and sip-and-dippers. The timing of the bids is
used as the defining attribute in the categorization of the bidder types. The evaluators
submit only one bid in the auction. The bid is placed either in the beginning or
towards the middle of the auction. The evaluator knows her valuation for the item, and
minimizes her effort cost from participating in the auction. The downside is that she
may end up paying too much (in a forward auction) or selling too cheap (in a reverse
auction). The opportunists are bargain-hunters. They enter the auction late and hope
that by bidding late they will leave little time for competitors to act. When the ending
rule is flexible, as in CombiAuction, the opportunist’s strategy is not as effective as in
an auction with a fixed closing rule. However, considering a wide variety of auctions
(and not a single isolated auction), the opportunist’s strategy of waiting to see how
fierce the competition in each auction is before participating, can be effective even
with flexible closing rule. Participants spend a lot of time bidding in the auction. They
start bidding early on, and continue bidding until the closing time approaches. The
sip-and-dippers participate in the beginning and at the end, but not in between. A
typical sip-and-dipper places two bids: one bid at the beginning to establish her
presence and to assess competition, and one bid that reveals their valuation when the
closing time approaches.
183
Shah et al. (2003) also identify four bidder strategies: evaluator strategy, skeptic strategy,
sniping strategy and unmasking strategy. Instead of using the timing of bids as the basis
of categorization, Shah et al. examine whether bidders bid above the required
increment (in a forward auction). Evaluators usually bid significantly higher than what
the increment suggests – in addition to bidding early in the auction. Skeptics bid from
the beginning to the end, as do the participants of Bapna et al. (2004), and they
typically increase their bid exactly by the required increment. The name “skeptic” is
derived from the fact that a proxy agent was available in the auctions Shah et al. (2003)
studied, but these bidders did not use it. They rather bid manually every time they
became outbid. The sniping strategy was equivalent to opportunist’s strategy of Bapna
et al. (2004). The unmasking strategy consisted of a series of bids submitted close to
each other. Shah et al. (2003) suspect that the purpose of such a strategy is to try to
reveal other bidders’ proxy bids – and hence the name. This strategy is specific to
auctions that enable proxy bidding, and thereby was not considered by Bapna et al.
(2004).
Based on previous literature, Puro (2009) develops a categorization of bidders’
strategies that combines the timing of bidding, the number of bids placed, and the size
of the bid decrement below the minimum required decrement. Puro studies online
people-to-people (P2P) auctions in which are conducted as reverse auctions. Puro
draws a distinction between strategies in which bidder places only one bid in the entire
auction and strategies with multiple bids. He identifies five single-bid strategies:
sniping, late bidding, opportunist, evaluator and portfolio bidding, and four multi-bid
strategies: all late, all skeptic, last bid late and stepped bidding. In the single-bid
strategies late bidding refers to bidding within the last 12 hours of the auction. Sniping
and opportunistic bidding are special cases of late bidding. Snipers place their single
bid within the last 30 minutes of the auction. Opportunists place their single bid before
the snipers, and always with a price which is close to the required decrement. Thus,
Puro’s definition of an opportunist is a little different from that of Bapna et al. (2004).
In other words, the opportunists try their luck with one bid which becomes the leading
bid when placed, but which maximizes the probability of being outbid. Evaluators
place their bid before the last day of the auction, and always with a decrement much
184
larger than what would have been required. Portfolio bidding is specific to the
Prosper.com auction site Puro uses in his study. It refers to the possibility for bidders to
specify their preferred products and valuations, and a bidding agent places
automatically a bid in an auction that meets the bidder’s criteria. In the multi-bid
strategies all skeptic refers to a strategy similar to Shah et al.’s (2003) skeptic strategy. All
late strategy is the multi-bid version of late bidding, that is, all bids are placed within
the last 12 hours of the auction. Last bid late refers to a strategy in which bidder places
at least one bid within the last 12 hours, but has already bid before that as well. Stepped
bidding is the multi-bid equivalent of evaluator strategy. In stepped bidding the bidder
places several bids early in the auction with decrements much lower than required.
All the studies reviewed above use data from real life auctions. One can then assume
that a significant portion of the participants have prior experience from online auction.
Also, all of the previous studies focus on single-item (single- or multiple-unit) auctions,
which are conceptually much simpler bidding environments than combinatorial
auctions. Thus, the impact of inexperience of the participants cannot be forgotten in
analyzing the auctions in this experiment.
When combinatorial bidding is involved, strategies need to be evaluated somewhat
differently. The timing of bids is still a relevant attribute, but when inactive bids are
allowed to enter the bid stream (as is the case with CombiAuction) the bid
decrement/increment is not directly applicable any more. Of course, when bids are
placed with the help of support tools, the decrement is clearly defined. However,
whenever bidder places a self-made bid, she does not know what price is required to
make the bid active. Thus, in those cases, one cannot really talk about deliberately
bidding below the required decrement. In combinatorial auction bidders rarely place
only one bid, so the first five strategies identified by Puro (2009) are not applicable.
Also, in combinatorial auctions the combination of items and item quantities become a
defining characteristic in bidding strategies.
12.2.2.2 Bidders’ Strategies in the Laboratory Experiment
I analyzed the auctions in order to identify distinct strategies using the categorizations
in existing literature as guidelines. The relevant strategies are the evaluator/stepped
185
bidding strategy of placing very competitive bids early on in the auction, the
opportunist/sniping/all late strategy of entering the auction very late, the
participant/skeptic strategy of bidding throughout the auction, and the sip-and-dip
strategy of bidding at the beginning and at the end. The identification of these
strategies is based mainly on the timing of bidding, and the competitiveness of the bid
(profit margin below required decrement). However, an interesting aspect in
combinatorial auctions is the content and number of the bids (what item
combinations, which quantities, and how many different combinations one bidder bids
for). My goal is to identify strategies for forming bids, as this has not been done in
literature.
Thus, in order to identify bidding strategies, I studied the following characteristics in
the bidders’ behavior: 1) what time bidders place their first bid, 2) how often they bid
and whether they were actively monitoring the auction around the closing time, 3)
what item combinations and quantities they bid for in self-made bids, and 4) what
prices they attach to the self-made bids. In addition, I analyzed how all of these
attributes of the bidders’ behavior changed from the A auction to the B auction. I used
my observations from the auctions and the bidders’ comments from the post-
experiment questionnaire to answer these questions.
Time of Placing First Bid
The time a bidder chooses to enter the auction is one common attribute used to
categorize bidders. Bidding early is part of the evaluator/stepped bidding strategy,
participant/skeptic strategy as well as the sip-and dip strategy. Entering the auction late,
on the other hand, is the essence of the opportunist/sniping/all late strategy. I defined
early bidders as those, who placed their bid within the first hour of the auction,
because these bidders are clearly eager to start bidding. In the A auctions, 16 bidders
(21.6%) qualified as early bidders. In B auctions, only 6 bidders (8.7%) can be
categorized as early bidders, and 5 of those also were early bidders in the A auctions.
Even if the definition of early bidders were extended to bidding within the first two
hours, the numbers would rise only to 17 (A auctions) and 11 (B auctions). I defined
the late bidders as those who submitted their first bid during the last two hours of the
186
auction. It is clear that they have deliberately bid late, because having had 21.5 hours
time to bid they cannot plead a busy schedule as a reason for not bidding. In A
auctions, 9 bidders (12.2%) were late bidders, and in B auctions 18 bidders (26.1%) bid
late. One explanation for this is that because bidders realize they need to monitor the
auction when the closing time approaches anyway, they can minimize their effort by
starting to bid first then. Also, the opportunist strategy is the best strategy in simpler
auctions, like the ones in eBay. Perhaps the bidders learned it from there. A cruder
categorization is to look how many bidders began bidding during first day, and how
many waited until the second day24. In A auctions, 42 bidders (56.8%) bid during the
first day, but only 33 bidders (47.8%) bid during the first day in the B auctions.
A big portion of the bidders changed the timing of their first bids from A auction to B
auction. Out of the 69 bidders 43 (62.3%) bid either at least two hours later or two
hours earlier in the B auction than in the A auction, and 21 bidders (30.4%) even
changed the day they started bidding. Of course the participants’ schedules affect when
they have time to log into the auction system and focus on bidding. However, it is also
very likely that due to the complexity of the auction, the bidders experiment with
different strategies in hopes of finding a good one. The fact that the outcome of the A
auction (whether a bidder was a winner or a loser) does not correlate with the decision
to bid at a different time in the A auction supports the latter explanation. Out of the
losers of the A auctions 64.3% and out of the winners 59.3% changed the timing of
their first bid by more than 2 hours. It is also possible, that some bidders’ strategy was to
wait until the support tools became active, and started bidding only after that. This
hypothesis cannot be verified though, because there is no way of knowing if a bidder
had visited the auction site, unless she also placed a bid.
Frequency and Timing of Bidding
By observing the time of entry into the auction alone, only the opportunists/snipers can
be identified. In order to identify the other strategies, the entire bidding pattern
(frequency and timing) needs to be studied.
24 The auctions started at 5pm during the first day and the scheduled closing was at 4:30pm the following day.
187
The frequency of bidding varied a lot. Naturally, if your initial bid is active until the
very last minute, you have no incentive of placing more bids. Also, I cannot know how
many times the bidders used the support tools in order to place a bid but found no
feasible suggestion and therefore did not place a bid. Therefore, seeming inactivity of a
bidder (no bids in the bid stream) does not mean she would not have participated in
the auction actively and visited the auction site often. This is where I used the answers
bidders gave in the questionnaire to get a better understanding of the bidders’
strategies.
The 9 late bidders/opportunists in A auctions were identified above. Late bidding was
defined as entering the auction 2h before scheduled closing or later. If the definition is
extended to include bidders who entered within 3h of closing, the number increases to
11 (14.9%). I defined participant/skeptic strategy as bidding soon after auction opened
on the first day and then in the morning and afternoon of the second day all the way
until closing. There were 9 bidders, who followed a pure participant strategy. In
addition, there were 8 bidders who exhibited participant behavior, but did not enter the
auction until after 9 pm on the first day (over 4h after auction began), and 2 bidders
who participated from early on, but failed to monitor the closing of the auction. Thus,
depending on how strictly one wants to define the participant/skeptic strategy, either
12.2% or 25.7% of bidders followed it. There was also a group of 7 bidders, who I call
“partial participants”, who started bidding on the morning of the second day and bid
until the end.
There were only 5 sip-and-dippers (6.8%), if defined as bidding within the first 3 hours
and then the 2 last hours of the auction. If this definition is extended to include
bidding later on the first day, and returning in the afternoon the following day (when
4h were left), the number of sip-and-dippers increases to 16 (21.6%) of bidders.
The behavior of only 5 bidders (6.8%) could be interpreted as that of an evaluator.
Characteristic of these bidders was that they placed only few bids either at the
beginning or middle of the auction, and at a price close to production costs. Also a
common characteristic is that they did not use the support tools to place the bids. The
small number of evaluators is not surprising, because it is difficult to know what
188
combinations to bid on if one does not monitor the auction and try out different
combinations. By bidding on only a few combinations – albeit with a low price – and
trusting that other bidders will then place complementing bids, the bidder takes a big
risk. Interestingly, though, four of these evaluators ended up winning (although one
actually made a loss with her bid). One explanation is that by placing a highly
competitive bid relatively early, their bid became the bid the QSM would use as a
complementing bid in most suggestions. However, the profits made by the winning
evaluators were much lower than the average profit among the winners.
The remaining 16 bidders did not follow any identifiable strategy. Some of them did
not monitor the auction until the end, some of them entered the auction rather late
but not late enough to be an opportunist, and some did not provide enough
information on their strategy in the questionnaire.
There were winning bidders among all the bidder categories. This is because strategy is
not the only determinant of winning; cost matters as does the content of the bids, and
the kind of bids the competitors have made.
The 18 late bidders/opportunists in the B auctions were identified already earlier. Just
as in A auctions, if the definition of late bidding is extended to 3 hours before closing,
the category is extended by two bidders to 20 (30.0%). In this second set of auctions,
only 5 (7.2%) of the bidders could be categorized as participants. If the category is
extended to those entering the auction later on the first day, 4 bidders can be added. If
also the bidder who started early but did not participate at the end is added, the total
number of bidders following the participant strategy is 10 (14.5%). This is clearly less
than in the A auction. Perhaps based on their experience in the A auctions the bidders
thought that the participant strategy is not the best one, and switched to other
strategies. Many bidders seemed to have learned the importance of the last hour of
bidding, which led them to reduce their effort prior to that. One indication of this is
that 6 bidders categorized as participants in the A auction were now using the late
bidding strategy. Another possibility is that the excitement of the game had worn off
already during the first auction, and the bidders did not want to spend so much time in
the second auction. One indication of this is that 5 bidders failed to bid in the second
189
auction altogether. The number of partial participants was 8 (11.2%), which is almost
the same as in the A auction.
There were 6 strictly sip-and-dipper bidders (8.7%) in the B auctions, and extending
the definition increased the number to 11 (15.9%). As was the case with participants,
also from the sip-and-dippers in the A auctions some bidders switched to late bidding.
However, also some bidders switched to the participant strategy. The number of
evaluators was again 5 (7.2%), and three of them were winners – although one of them
had miscalculated their cost and actually won with an unprofitable bid.
The rest of the bidders (20 = 29.0%) did not fit into these categories, or they had not
provided enough information on their bidding strategy in the questionnaire. Like in
the A auctions, it is difficult to say, which strategy would have been the best. Bidders
from all categories were among the winners.
Table 34 summarizes the bidder strategies observed in the all the experiment auctions.
Table 34 Frequencies of bidder strategies in the experiment auctions
Late Bidders
Participants Partial Participants
Sip & Dippers
Evaluators Other
A Auctions*
11 (14.9%) 19 (25.7%) 7 (9.5%) 16 (21.6%) 5 (6.8%) 16 (21.6%)
B Auctions*
20 (30.0%) 10 (14.5%) 8 (11.2%) 11 (15.9%) 5 (7.2%) 20 (29.0%)
*There were 74 bidders in the A auctions and 69 in the B auctions.
Of course, one must keep in mind that strategy is not the only thing determining the
frequency and timing of bidders’ bids. The bidders (students) have busy schedules, and
perhaps cannot participate in the auctions as much as they would like to. Also, some
students are more motivated than others to make time for participation.
Item Combinations and Quantities
One interesting aspect of bidding behavior in combinatorial auctions is the kind of bids
the bidders place, namely which items and what quantities the bidders bid for. In
CombiAuction the availability of the QSM muddles the data somewhat, because the
QSM gives bidders suggestions on item combinations and quantities. However, at least
in the beginning of the auctions, when the support tools were not available, it was
possible to observe the contents of the bidders’ self-made bids.
190
One division can be made between proactive and reactive bidders. Proactive bidders
place a wide array of different item combinations - “hooks in the water”, as one bidder
put it – so that other bidders’ bids could team up with them. A few bidders submitted
bids for all 31 item combinations (at maximum capacity). Reactive bidders resorted to
the support tools, and placed bids based on the suggestions from the tools. As many as
13 bidders in both auctions report having placed bids only with the help of support
tools, once they became available. Some bidders report using mainly the support tools
to submit new bids, but placed self-made bids when the support tools suggested only
unprofitable bids. The majority of the bidders were somewhere in between the two
extremes; they placed self-made bids but also used the support tools.
Another division can be made between intelligent and random bidders. Intelligent
bidders placed bids on particular combinations for justified reasons. In total 16 bidders
said they tried to bid for combinations for which they had the lowest cost. It is not
trivial, how to determine the combinations with the largest cost advantage in a
combinatorial auction, when you do not have much information on your competitors’
cost. Because in the set-up of this experiment, all items were treated equally (costs
drawn from the same distributions), many bidders cleverly calculated per-unit average
costs for bundles. If the items in the bundle had very different variable and fixed costs,
such per-unit calculations across items could not have been done. In addition, 5
bidders said they tried to bid for large quantities to minimize the average cost. Where
the intelligent bidders placed only a few well thought bids, the random bidders did not
seem to have any logic in which items and quantities they placed. At the extreme, a
random bidder would place bids for every item combination, and several bids with the
same items, but different quantities.
The most popular bid was to place a bid containing all five items at full capacity,
which was to be expected. Bidding for full capacity seemed to be the most popular
choice in item quantities overall, which makes sense because of the economies of
scale. In addition to bidding for all five items, bidders bid for various subcombinations
of the full capacity bid (including bids for single items). We anticipated such behavior
already when designing the “less support” benchmark cases in the second simulation
study. Also the bid for all items, but with ½ of maximum capacity as quantities (which
191
we also used in the second simulation study) was submitted in some of the equal
capacities (A) auctions, but not very often. In the unequal capacities auction no such
bids were placed. The bidders turned out to be much more creative than we had
anticipated. A vast array of different bids (other than full capacity) was placed. A
common denominator in the choice of bid quantities was to use fractions of the total
demand (e.g. 100, 100, 100, 100, 100)25 or (200, 200, 200, 200, 200). The bidders had a
tendency to bid for equal quantities of all items also in the unequal capacities auction.
Another popular choice was to bid for quantities that might complement other bids
(e.g. 75, 75, 75, 75, 0) in the unequal capacities auctions where some bidders had
capacity limits of 225. The bidders did not restrict themselves to even quantities across
items; bids like (300, 100, 200, 300, 200) and (0, 300, 150, 150, 0) were placed also in
the equal capacities (A) auctions, but were more common in the unequal capacities
(B) auctions. The bidders seemed to have a tendency to place “desperate” bids (= bids
with virtually zero profit) on small quantities of just one or two items when they could
not find any active bids in the auction. Evidently they did not understand that due to
economies of scope and scale the smaller bids had an even lower probability of
becoming active.
In 8 of the 13 A auctions and 2 of the 11 B auctions some bidders placed bids which
did not follow any of the bid creating logics described above. Bidders could place bids
like (134, 200, 85, 93, 240), (220, 220, 220, 220, 220), and even (0, 0, 0, 1, 0). The fact
that there were fewer odd bids in the B auctions leads me to believe that inexperience
and frustration caused some of the bidders to bid completely randomly. Also, these odd
bids were placed by bidders, who had difficulties in grasping the auction concepts
(several of them placed bids that exceeded their capacity or resulted in a loss for them),
or who placed dozens of bids during the auction.
Bid Prices
Bidders followed very different pricing strategies. Five distinct strategies could be
defined: opportunists, evaluators, satisficers, support users, and desperate bidders.
Opportunists tried their luck by submitting bids with very high margins (over 100%,
25 I have left the bid prices out from the bids because they are irrelevant in this discussion.
192
sometimes even 1000%) at the beginning of the auctions. As the auction proceeded,
they were forced to lower their prices drastically in order to stay competitive. The
highest initial margins were observed only in the A auctions; bidders soon learned that
such bids would not become active. However, some bidders did manage maintain
profit margins as high as 20% in some of he B auctions. Evaluators were more
concerned about winning than making a large profit. Moreover, they did not want to
spend much time monitoring the auction and repeatedly submitting bids. Thus, they
placed a few bids with very low margins before the heated competition around the
closing of the auction. Usually the evaluators also study their cost function to identify
the combinations where they should have a cost advantage. Satisficers have a profit
target in mind (either profit margin or amount of euros) when they set out to bid. They
place bids – both self-made and obtained with the help of a support tool – in which the
target is achieved. Sometimes they even place bids with a lower price than suggested by
the support tools, if the suggestion contains a higher profit. The support users put the
least effort into the bidding process in the sense that they did not make any
preparations prior to the auction. They simply used the support tools and evaluated the
suggestions. This way their time cost from participation was smaller than for other
bidders. Of course, sometimes they needed to put a lot of effort in the evaluation
process – e.g. when the shortlist offered by the QSM was long – but it was rather
mechanical a task.
The fifth strategy, desperate bidding, is a strategy any of the bidders can switch to when
the support tools are not helpful, and the bidder does not have any active bids. Usually
this happens near the closing time, but it can happen at any point in the auction when
a bidder desperately wants to become active. Desperate bidding entails placing bids on
seemingly random combinations of items and with very low profit margins (even lower
than in the evaluator’s bids). Usually the bidder places several such bids within a very
short time period.
193
12.2.3 Other Observations on Bidders’ Behavior and the CombiAuction
System
The last objective of the laboratory experiment was to study the usability of the
CombiAuction website. This means studying the bidders’ perceptions of the site and
the support tools. Also, studying how well the bidders learned to use the auction system
and how well they understood the auction concepts is relevant, as it can validate or
undermine the results presented in the two previous sections. There is little point in
studying the outcomes of the auctions or the bidders’ strategies, if the bidders had no
idea of what they were supposed to do. Also, studying the bidders’ perceptions and
understanding of the auction system has implications on the organization of future
auctions.
Usability of the User Interface and Support Tools
The participants were asked to rate how easy the user interface of the CombiAuction
was to use, and how helpful the price and quantity support tools were (see Appendix 5).
The average rating for the easiness of use (scale 1-5, 5 being the easiest) was 4.23
(standard deviation 0.80). The main complaint from the participants was that towards
the end, when all bidders were logged in at the same time, the auction system became
slow.
The average score for the helpfulness of price support was 3.81 (st. dev. 1.20) and for
the helpfulness of quantity support 4.05 (st. dev. 1.10) on a scale 1-5 with 5 being most
helpful. Most participants (45 out of 66) rated the price support and quantity support
tools equally (un)helpful, which I had not anticipated. I imagined the bidders would
find the quantity support tool more helpful. Moreover, a few participants (6 out of 66)
even rated the price support tool more helpful than the quantity support. It could be
that the combination of price and quantity support is much better than either of them
alone. The QSM provides the bidders with good item combinations and they can then
use the price support to keep the bids active as the auction proceeds.
Another explanation is that bidders do not care about the efficiency of the auctions.
Knowing that they were not supposed to win does not make them feel any better about
losing the auction – and without knowledge of other bidders’ cost functions they
194
cannot even know that they were not supposed to win. When they fail to find profitable
bids they are equally unsatisfied with both support tools. This would also explain why 8
bidders answered the question on the helpfulness of the quantity support either “I
somewhat disagree” (= 2) or “I totally disagree” (= 1). If at the end of the auction a
bidder cannot find profitable bids from the shortlist, she is frustrated. It does not matter
to her what the reason for that is. If the auction has ended in an efficient allocation
with the total cost so low that others cannot afford to bring it down by the required
decrement, the auction owner is pleased. The losing bidders, however, do not know
what the reason for unprofitable suggestions is; they just see that the auction is closing
and they are not among the winners.
Observations on Bidder Behavior and Understanding
The participants claimed that they understood the rules of the auctions and what their
goal was. The average ratings for these questions in the questionnaire were 4.45 and
4.49 respectively (st. dev. 0.93 and 0.68) on a scale 1-5 with 5 indicating they
understood well. However, the bidders made a lot of mistakes in the auctions. At least
14 bidders placed bids that exceeded their capacities, and at least 19 bidders submitted
bids with a negative profit. Some of these mistakes were simply the result of
carelessness; bidders forgot one zero from the price, they accidentally clicked on an
undesirable bid in the support tools, had a mistake in their excel calculations, or made
a typo when submitting a self-made bid (even though a confirmation window opens
every time a bid is submitted). Several mistakes took place when the auction was about
to end. Bidders were under pressure to act quickly in order to submit their bid before
the auction closed, which led to carelessness. Also, it is possible that some bidders had
so strong a desire to win, they deliberately made unprofitable bids. However, I did not
find any indication of this in the questionnaires or in the emails I exchanged with some
students.
Many times the bidder noticed the mistakes themselves and contacted me asking if I
could delete the bids. However, among the winners there were 9 bidders, who won the
auction with an unprofitable bid. There were also other signs of bidders not totally
understanding how a combinatorial auction and the support tools worked. Bidders
195
sometimes interpreted the inactive status to be the result of too high a price, when in
fact it was due to lack of other bids in the auction. This led the bidders to unnecessarily
lower the bid prices already at the beginning of the auction, when there were no other
bidders present. A few bidders were also bewildered when the support tools were not
available, and contacted me asking why they were not available. Also, in the participant
questionnaire, three students answered that they did not understand how the QSM
worked, and one even wrote that he “didn’t understand the tools so good [well] that
[he] could rely on them”. Clearly, the briefing session was not enough for some of the
participants to understand the concept of a combinatorial auction and of the support
tools. However, these were isolated instances, and they have not affected my analysis of
the auction outcomes or the analysis of the bidders’ strategies.
The challenge of combinatorial auctions is that participating in them requires an
understanding of the combinatorial aspects of the auction. In order to understand the
winner determination and the quantity support, the bidders should have knowledge of
the basic concepts of optimization. If not, the auction appears to them as a black box,
and they cannot discern the link between their actions and the auction outcomes. Also,
not understanding how the support tools work, bidders may not use them – or expect
too much from them and become frustrated. What this experiment clearly shows is that
briefing the participants in advance is a crucial phase in holding combinatorial
auctions.
It is relatively easy to understand that cost advantages matter – and several participants
in the experiments had understood this and tried to act accordingly. However, it is not
trivial to understand that cost advantage is ultimately a relative concept: it is enough to
have relatively low costs on items for which there are good complements. Thus, other
bidders’ costs affect what is the best bid for you. Unfortunately, even if a bidder
understands this, there is not much she can do, because she does no know the other
bidders’ cost functions. All the bidders can try to do is to place some self-made bids on
combinations for which her costs are low, and to use the support tools in order to find
combinations to team up with other bids. This is not sufficient to guarantee a winning
bid in the auction – but not using any support makes winning more difficult, or at least
decreases the profits in the winning bids.
196
VII CONCLUSIONS
Combinatorial auctions have become an interesting subject of research. The literature
focused on different aspects of combinatorial auctions has increased significantly
during the past decade. There have also been some applications of combinatorial
auctions into practice. However, there are still many ways combinatorial auctions
could be improved to make them easier to use and applicable to a wider array
situations.
The underlying problem with combinatorial auctions is that they are complex in many
ways. Not only are combinatorial auctions computationally difficult, but they are also
challenging environments for bidders. In this thesis I have introduced the puzzle
problem. The puzzle problem refers to the situation in which bidders should
coordinate the items and item quantities they bid for (in addition to price), in order to
overcome the current provisional winners. Usually all communication among bidders
is forbidden, as bid takers try to prevent collusion among bidders. With collusion I
mean bidders’ attempts to benefit at the expense of the bid taker. Also, reverse auctions
are often sealed-bid (or semi-sealed-bid) auctions meaning that the bidders do not
know the contents of their competitors bids. The problem facing a bidder in such an
auction can be compared to the task of trying to complete a puzzle without knowing
the size and shape of the missing piece. Hence the name “puzzle problem.”
In our research project we developed a bidder support tool called the Quantity Support
Mechanism (QSM) to help bidders overcome the puzzle problem in continuous, semi-
sealed-bid combinatorial auctions. The QSM can be used equally in forward and
reverse auctions, but in this thesis I have presented it in the reverse auction setting. At
the heart of the QSM there is an IP problem (QSP), which maximizes the
approximated profit (difference between the price in the new bid and the approximated
cost of producing the suggested item quantities) of a currently non-winning bidder
subject to the total cost to the buyer decreasing by a required decrement and the
buyer’s demand being fulfilled. The bidders’ profit is approximated with a linear cost
function. The estimates for the per-unit costs for each item are obtained from the dual
prices of the linear relaxation on the Winner Determination Problem. Because the
197
approximation is anonymous (= each bidder is assumed to have the same costs), and
because the linear approximation may be far off from the true cost function (which
presumably exhibits economies of scope since it is a combinatorial auction), a shortlist
of alternative bid suggestions is created. The shortlist is created by forcing non-basic
variables into the basis (i.e. forcing inactive bids among the provisional winners) and by
forcing different item combinations to zero, each at a time. The shortlist is offered to
the bidder, who then decides which bid to submit, if any.
We ran simulations to test the QSM. The simulations showed that the QSM helped
solve the puzzle problem. The first simulation study showed that the QSM found good
suggestions for the bidders, and that it produced better suggestions than a QSM which
used random cost parameters instead of the dual prices. The second simulation study
showed that the QSM improved the efficiency of the auction outcomes compared to
the situation in which only price support was available.
The QSM was implemented in an online auction system called the CombiAuction.
The auction system is designed for continuous combinatorial auctions, but it is up to
the auction owner to decide whether the auction is a reverse or a forward auction and
whether it is a semi-sealed-bid or an open-cry auction. The auction owner can also
decide, whether the bidders have access to the QSM and the price support. The
CombiAuction system was tested in an experiment with human subjects. The objective
of the experiment was threefold. Firstly, the objective was to study the usability of the
user interface of the CombiAuction system as well as the QSM. Secondly, I wanted to
study the bidders’ behavior and strategies in the auction. The third objective was to
compare the outcomes of the experiment auctions to those of the simulations.
From the equal capacities auctions about half ended in an efficient (or pseudo-
efficient) allocation, as had the simulated auction. The rest fell prey to bidders who
made unprofitable bids or bidders who could not monitor the auction at the end. In
addition, one auction fell prey to the threshold problem. Also in several of the unequal
capacities auctions many bidders won with an unprofitable bid, which distorted the
final allocation. Out of those auctions in which bidders did not win with unprofitable
bids, one third ended in the efficient allocation. This is a good result seeing that the
198
corresponding simulated auction ended in a very inefficient allocation (the production
cost of the winning allocation was 9.2% above efficient cost). The good outcomes of
the laboratory experiment auctions compared to the simulated auctions also helped
validate the simulation results in general. Whenever simulating human behavior, there
is the risk that by making simplifying assumptions one creates a more favorable setting,
which leads to too good results. However, since the auctions with real bidders ended
up in similar or better outcomes than the simulated auctions, the assumptions we
made in the simulations seem to have been reasonable.
In the literature, there are several classifications of bidder strategies in online auctions.
Using these categorizations as guidelines I could identify four bidding strategies: late
bidders (opportunists, snipers), participants (skeptics), sip-and-dippers, and evaluators.
However, a significant portion of the bidders did not fit into these categories, nor did
they form new categories. This was partially due to not having enough information on
their strategy, and partly due to not being able to detect any patterns from their bidding
behavior. The above-mentioned strategies refer to the timing of bidding, and also the
size of the mark-up in the bids (with evaluators). The strategies bidders use to form
their combinatorial bids have not been studied in the literature. Based on the bid data I
formed two categorizations: proactive vs. reactive bidders, and intelligent vs. random
bidders. Proactive bidders place a lot of bids on different combinations for other bidders
to team up with, whereas reactive bidders mainly use the support tools and try to find
bids that complement existing bids. When creating self-made bids, intelligent bidders
choose the item combinations based on their cost function. They try to place bids
where they think they may have a cost advantage. Intelligent bidders also bid at full
capacity in order to benefit from the economies of scale. Random bidders bid on many
different combinations with no seeming logic, and they typically bid on more
combinations than the intelligent bidders. Random bidders also often choose item
quantities below their maximum capacity. Also intelligent bidders sometimes choose
quantities below their maximum capacity when attempting to place bids that they
expect to team up well with other bids or sum up easily to total demand. For example,
if the bidder’s capacity is 225 units for a particular item, and the demand is 600 units,
199
the intelligent bidder may choose to bid for 200 units. All in all, the bidders proved to
be more creative when creating self-made bids than I had anticipated.
Based on my observations and feedback given by the participants the interface seemed
to be good and easy to use. The only complaints were related to the technical
difficulties encountered during the last session (B2-auctions). However, the bidders
made some mistakes: they placed unprofitable bids and bids that exceeded their
capacity. We should redesign the interface to minimize the possibility of making such
mistakes. For instance, whenever selecting a suggested bid from the shortlist, a
confirmation window should pop up. Also, making it easier to export data from the
auction system to Excel and to import it back might reduce mistakes. Based on my
observations, some bidders did not quite grasp the concept of a combinatorial auction;
perhaps a more thorough briefing session combined with a quiz would be called for.
The second simulation study revealed that the QSM did not always guide the auctions
to the efficient allocation. Moreover, whenever the efficient allocation consisted of
more than two bids, the efficient allocation was hardly ever found. The QSM suffers
from the extended puzzle problem, which is in way an extension of the threshold
problem well recognized in the literature. The QSM is good at finding the last missing
piece to the puzzle, but unless the other pieces are the ones from the efficient
allocation, it will not find the efficient bids. This is because it relies on the existing bids
to find good complements for the new bid. Just as in the threshold problem it is useless
for one bidder to decrease bid price alone, it is useless for the bidder to place an
efficient bid unless its complements are present. It would not become active (a
provisional winner), and thereby the QSM will not find it, because the QSM is
designed to find bids that will become active immediately upon submission.
In order to help bidders overcome the extended puzzle problem we designed a new
support tool, the Group Support Mechanism (GSM). The GSM is based on the same
principles as the QSM: it aims at maximizing the profit of the incoming bidders while
decreasing the total cost to the buyer and fulfilling the total demand. The biggest
difference is that it suggests bids for several bidders at a time, and together the group of
bidders would become provisional winners. Another major difference is that in the
200
GSM the cost function approximations are customized for each bidder based on the
information obtainable from their bids. Based on preliminary tests the GSM seems to
improve the efficiency of auction beyond what the QSM could do. This result is
logical, since the GSM is essentially an extension of the QSM. As a special case the
GSM could also suggest a bid for only one bidder – like the QSM would – if it was the
optimal course of action. The weakness of the GSM is that whenever one of the
bidders does not submit her bid suggestion, none of the bidders become active. This
slows down the convergence of the auction and can cause frustration in the bidders.
Implications on Applying Combinatorial Auctions to Practice
The simulation studies already indicated that support is beneficial for the auction
owner in semi-sealed-bid auctions. The laboratory experiment strengthened this
impression. Moreover, the QSM seemed to produce better (= more efficient) outcomes
for the unequal capacities auctions in the laboratory experiments than was expected
based on the simulations. However, bidders made a lot of mistakes in the auctions. On
the one hand this calls for training and experience, but also challenges the design of
the user interface. When submitting a self-made bid in CombiAuction, a confirmation
window always opens. However, no such window opened if placing a bid through the
support tools. In both price and quantity support the bid was immediately submitted, if
the bidder clicked on the “Submit this bid” link. Consequently, bidders made more
mistakes with bids from the support tools than self-made bids. It is also very important
that all numbers – especially large ones – are easy and quick to read. The bidder needs
to be able to differentiate between a million and hundred thousand at a glance without
having to count the zeros. The experiment showed that bidders made a lot of mistakes
towards the end of the auctions when they were in a hurry. Therefore, the closing rule
should not be fixed, but flexible (as it was in the experiment), and the extension time
should be long enough that bidders can choose their course of action and estimate the
profitability of all the bid suggestions. The 10 minutes used in the experiment may not
have been long enough for that. The bidders’ behavior in the experiment indicated that
with more experience bidders shift their bidding closer to the end of the auction.
Hence the auction owner should anticipate heavy bidding activity during the last hour
201
of the auction. Naturally, this sets some requirements for the robustness of the system
as well; it needs to be able to handle heavy traffic.
The laboratory experiment showed that combinatorial auctions are indeed difficult
environments for bidders to bid in – especially for inexperienced bidders. One of the
reasons is that a combinatorial auction does not advance linearly: your more recent
bids can become inactive and older ones active. When the bidders did not understand
the logic of winner determination and auction progression, it resulted in confusion and
frustration. The participants felt that the support tools were easy to use, but some of the
bidders decided not to use them because they did not understand how they produced
the bid suggestions. Clearly, the bidders need to be briefed thoroughly on the
intricacies of combinatorial auctions, and the support tools. What is sufficient
information and how it is best conveyed is still an open issue. The need for training
effectively rules out the possibility of organizing online combinatorial auctions open
for everyone with access to the Internet. Fortunately, combinatorial auctions have
many applications in the B2B markets, in which the bidding is done by professional
sellers/buyers, who can be trained and who can accumulate the required experience.
Future Research
In this thesis I have brought up new insights into the challenges of combinatorial
auctions. I have also described the tools we have developed to tackle these challenges.
However, there are still many issues that need further studying.
First of all, the GSM should be developed further. The estimation of the bidders’ cost
functions, which is a crucial element in the GSM, should be studied further. Different
alternatives for the estimation and updating procedure will be developed and
compared with each other. Also the possibility of using the “pseudo dual prices” used
in many existing combinatorial auction mechanisms could be explored. The current
form of the cost function used in the Cost Estimation Problem has too many
parameters to be estimated from the limited bid information. Hence, the first step will
be to explore alternative, simpler forms for the cost function, which would still exhibit
economies of scope. The GSM should also be tested, first through simulations and
then in laboratory experiments with human subjects. Through simulations it is possible
202
to compare different configurations of the GSM. The laboratory experiments on the
other hand provide knowledge of how real bidders perceive and understand the GSM,
and how convenient it is to use.
Secondly, the effect of bidders’ learning and experience in combinatorial auctions on
strategies and auction outcomes should be studied. In the laboratory experiment
described in this thesis the bidders did not have prior experience of combinatorial
auctions. Everybody participated in two auctions, but one auction hardly gives enough
experience in such a complex bidding environment so that one could call the bidders
experienced in the second auction. However, because combinatorial auctions are
complex bidding environments, I expect experience to have a significant effect on
bidders’ strategies and thereby potentially on the auction outcomes. Already there were
some indications of learning taking place. For instance, some bidders learned wait
until close to the end to bid.
The QSM should also be fine-tuned before the next experiments. At least the length of
the shortlist should be restricted to ease the bidders’ burden of evaluating bid
suggestions. It is not trivial, what is the optimal way of shortening the shortlist, and
hence some alternative methods should be tested. Also, the system should be
redesigned to record more information on bidders’ strategies. For example, it would be
good to know afterwards, when the bidders were logged in the system, and what the
contents of the shortlists offered by the QSM were. It would also be interesting to
compare the QSM and GSM in a laboratory setting to get a better understanding of
what kind of allocations they lead to with human users, and which tool the bidders
prefer.
203
References
Adomavicius, G., and Gupta, A. (2005), Toward Comprehensive Real-Time Bidder Support in Iterative Combinatorial Auctions, Information Systems Research 16: 169–185.
An, N., Elmaghraby, W., and Keskinocak, P. (2005), Bidding Strategies and Their Impact on Revenues in Combinatorial Auctions, Journal of Revenue and Pricing Management 3(4): 337–357.
Andersson, A., Tenhunen, M., and Ygge, F. (2000), Integer Programming for Combinatorial Auction Winner Determination, in: Proceedings of the Fourth International Conference on Multi-Agent Systems (ICMAS), Boston, MA.
Ausubel, L.M. (2004), An Efficient Ascending-Bid Auction for Multiple Objects, The American Economic Review 94: 1452–1475.
Ausubel, L.M., Cramton, P., and Milgrom, P. (2006), The Clock-Proxy Auction: A Practical Combinatorial Auction Design, Chapter 5 in: P. Cramton, Y. Shoham, and R. Steinberg (eds.), Combinatorial Auctions, MIT Press, 115–138.
Ausubel, L.M, and Milgrom, P. (2002), Ascending Auctions with Package Bidding, Frontiers of Theoretical Economics 1: 1–42.
Ausubel, L.M., and Milgrom, P. (2006), The Lovely but Lonely Vickrey Auction, Chapter 1 in: P. Cramton, Y. Shoham, and R. Steinberg (eds.), Combinatorial Auctions, MIT Press, 17–40.
Bajari, P., and Hortaçsu, A. (2003), The Winner’s Curse, Reserve Prices, and Endogenous Entry: Empirical Insights from eBay Auctions, RAND Journal of Economics 34(2): 329–355.
Banks, J.S., Ledyard, J.O., and Porter, D.P. (1989), Allocating Uncertain and Unresponsive Resources: An Experimental Approach, RAND Journal of Economics 20: 1–25.
Banks, J., Olson, M., Porter, D., Rassenti, S., and Smith, V. (2003), Theory, Experiment and the Federal Communications Commission Spectrum Auctions”, Journal of Economic Behavior and Organization 51: 303–350.
Bapna, R., Goes, P., and Gupta, A. (2000), A Theoretical and Empirical Investigation of Multi-Item On-line Auctions, Information Technology and Management 1(1–2): 1–23.
Bapna, R., Goes, P., and Gupta, A. (2003), Analysis of Design of Business-to-Consumer Online Auctions, Management Science 49(1): 85–101.
Bapna, R., Goes, P., Gupta, A., and Jin, Y. (2004), User Heterogeneity and its Impact on Electronic Auction Market Design: An Empirical Exploration, MIS Quarterly 28(1): 21–43.
Bazerman, M.H. and Samuelson, W.F. (1983), I Won the Auction but Don’t Want the Prize, Journal of Conflict Resolution 27: 618–634.
204
Beil, D.R., and Wein, L.M. (2003), An Inverse-Optimization-Based Auction for Multiattribute RFQs, Mangement Science 49(11): 1529–1545.
Bichler, M. (2000), An Experimental Analysis of Multi-Attribute Auctions”, Decision Support Systems 29(3): 249–268.
Bichler, M. (2001), The Future of e-markets: Multidimensional Market Mechanisms, Cambridge University Press, Cambridge, UK.
Binmore, K. and Klemperer, P. (2002), The Biggest Auction Ever: The Sale of the British 3G Telecom Licenses, The Economic Journal 112: C74–C96.
Borcherding, K., Eppel, T., and von Winterfeldt, D. (1991), Comparison of Weighting Judgments in Multiattribute Utility Measurement, Management Science (37): 1603–1619.
Branco, F. (1997), The Design of Multidimensional Auctions, Rand Journal of Economics (28): 63–81.
Cantillon, E., and Pesendorfer, M. (2006), Auctioning bus routes: the London experience, Chapter 23 in: P. Cramton, Y. Shoham, and R. Steinberg (eds.), Combinatorial Auctions, MIT Press, 573–592.
Capen, E., Clapp, R., and Campbell, W. (1971), Competitive Bidding in High Risk Situations, Journal of Petroleum Technology 23: 641–653.
Caplice, C. (2007), Electoronic Markets for Truckload Transportation, Production and Operations Management 16(4): 423–436.
Caplice, C., and Sheffi, Y. (2006), Combinatorial auctions for transportation services, Chapter 21 in: P. Cramton, Y. Shoham, and R. Steinberg (eds.), Combinatorial Auctions, MIT Press, 539–572.
Che, Y.-K. (1993), Design Competition through Multidimensional Auctions, RAND Journal of Economics 24: 668–679.
Chen-Ritzo, C.-H., Harrison, T.P., Kwasnica, A.M., and Thomas, D.J. (2005), Better, Faster, Cheaper: An Experimental Analysis of a Multiattribute Reverse Auction Mechanism with restricted Information Feedback, Management Science 51: 1753–1762.
Cho, Y. (2003), Economic Efficiency of Multi-Product Structure: The Evidence from Korean Housebuilding Firms, Journal of Housing Economics 12: 337–355.
Clarke, E.H. (1971), Multipart Pricing of Public Goods, Public Choice 11: 17–33.
Conen, W. and Sandholm, T. (2001), Minimal Preference Elicitation in Combinatorial Auctions, Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), Workshop on Economic Agents, Models, and Mechanisms, Seattle, WA, August 6th (2001).
Coppinger, V.M., Smith, V.L., and Titus, J.A. (1980), Incentives and Behavior in English, Dutch and Sealed-Bid Auctions, Economic Inquiry 18(1): 1–22.
Cox, J.C., Smith, V.L., and Walker, J.M. (1988), Theory and Individual Behavior of First-Price Auctions, Journal of Risk and Uncertainty 1: 61–99.
205
Crescenzi, P., and Kann, V. (2006), A Compendium of NP Optimization Problems, (http://www.nada.kth.se/~viggo/wwwcompendium/node142.html, accessed Nov. 16th, 2006.
Cripps, M., and Ireland, N. (1994), The Design of Auctions and Tenders with Quality Thresholds: The Symmetric Case, The Economic Journal 104(423): 316–326.
Day, R.W., and Milgrom, P. (2008), Core-selecting package auctions, International Journal of Game Theory 36: 393–407.
Day, R.W., and Raghavan, S. (2007), Fair Payments for Efficient Allocation in Public Sector Combinatorial Auctions, Management Science 53(9): 1389–1406.
Day, R.W., and Raghavan, S. (2008), A combinatorial procurement auction featuring bundle price revelation without free-riding, Decision Support Systems 44: 621–640.
De Vries, S., Schummer, J., and Vohra, R. (2007), On ascending Vickrey auctions for heterogeneous objects, Journal of Economic Theory 132: 95–118.
De Vries, S., and Vohra, R. (2003), Combinatorial Auctions: A Survey, INFORMS Journal on Computing 15: 284–309.
Dunford, M., Hoffman, K., Menon, D., Sultana, R., and Wilson, T. (2004), Testing Linear Pricing Algorithms for use in Ascending Combinatorial Auctions, working paper, Department of Systems Engineering and Operations Research, George Mason University, Fairfax, VA.
Engelbrecht-Wiggans, R. (1988), Revenue Equivalence in Multiple Object Auctions, Economics Letters 26(1): 15–19.
Engelbrecht-Wiggans, R. (2001), The Effect of Entry and Information Cost on Oral versus Sealed-Bid Auctions, Economics Letters 70: 195–202.
Epstein, R., Henriques, L., Catalan, J., Weintraub, G.Y., and Martinez, C. (2002), A Combinatorial Auction Improves School Meals in Chile, Interfaces 32(6): 1–14.
Ervasti, V., and Leskelä, R.-L. (2009), Allocative Efficiency in Simulated Multiple-Unit Combinatorial Auctions with Quantity Support, forthcoming in European Journal of Operational Research.
FCC, The Federal Communications Commission (2000), Comment Sought on Modifying the Simultaneous Multiple Round Auction Design to Allow Combinatorial (Package) Bidding, Public Notice DA 00-1075, http://wireless.fcc.gov/auctions/31/releases/da001075.pdf, accessed Nov 30th, 2008.
FCC, The Federal Communications Commission (2007), Notice and Filing Requirements, Minimum Opening Bids, Reserve Prices, Upfront Payments, and Other Procedures for Auctions 73 and 76, Public Notice DA-4171, http://fjallfoss.fcc.gov/edocs_public/attachmatch/DA-07-4171A1.pdf, accessed Oct. 17, 2007.
FCC, The Federal Communications Commission (2008), Public Notice DA 08-595, Attachment A: 62-63. http://hraunfoss.fcc.gov/edocs_public/attachmatch/DA-08-595A2.pdf, accessed Oct. 1, 2008.
206
Friedman, L. (1956), A Competitive Bidding Strategy, Operations Research 4(1): 104–112.
Fujishima, Y., Leyton-Brown, K., and Shoham, Y. (1999), Taming the Computational Complexitiy of Combinatorial Auctions: Optimal and Approximate Approaches, Proceedings of IJCAI -99, Stockholm, Sweden: 548–553.
Gallien, J, and Wein, L.M. (2005), A Smart Market for Industrial Procurement with Capacity Constraints”, Management Science 51: 76–91.
Garvin, S. and Kagel, J. (1994), Learning in Common Value Auctions, Journal of Economic Behavior and Organization 25: 351–372.
Goeree, J.K., and Offerman, T. (2002), Efficiency in Auctions with Private and Common Values: An Experimental Study, The American Economic Review 92(3): 625–643.
Gonen, R., and Lehmann, D. (2000), Optimal Solutions for Multi-Unit Combinatorial Auctions: Branch and Bound Heuristics, in Second ACM Conference on Electronic Commerce (EC-00), Minneapolis, Minnesota, October 2000.
Groves, T. (1973), “Incentives in Teams”, Econometrica 41: 617–631.
Guo, Y., Lim, A., Rodrigues, B., and Tang, J. (2006), Using a Lagrangian Heuristic for a Combinatorial Auction Problem, International Journal on Artificial Intelligence Tools 15(3): 481–489.
Guttman, R.H., and Maes, P. (1998), Cooperative vs. Competitive Multi-agent Negotiations in Retail Electronic Commerce, MIT Media Lab Paper, in Proceedings of the Second International Workshop on Cooperative Information Agents, Paris, July 1998.
Günlük, O., Lászlo, L., and de Vries, S. (2005), A Branch-and-Price Algorithm and New Test Problems for Spectrum Auctions, Management Science 51(3): 391–406.
Hoffman, K., Menon, D., and van den Heever, S. (2004), A Bidder Aid Tool for Dynamic Package Creation in the FCC Spectrum Auction, working paper, Dept. of Systems Engineering and Operations Research, George Mason University.
Hoffman, K., Menon, D., and van den Heever, S., and Wilson, T. (2006), Observations and Near-Direct Implementation of the Ascending Proxy Auction, Chapter 17 in: P. Cramton, Y. Shoham, and R. Steinberg (eds.), Combinatorial Auctions, MIT Press, 415–450.
Hohner, G., Rich, J., Ng, E., Reid, G., Davenport, A.J., Kalaganam, J.R., Lee, H.S., and An, C. (2003), Combinatorial and Quantity-Discount Procurement Auctions Benefit Mars, Incorporated and Its Suppliers, Interfaces 33: 23–35.
Isaac, R.M., and James, D. (2000), Robustness of the Incentive Compatible Combinatorial Auction, Experimental Economics, 3: 31–53.
Jap, S.D. (2003), An Exploratory Study of the Introduction of Online Reverse Auctions, Journal of Marketing 67: 96–107.
Jones, J.L., and Koehler, G.J. (2002), Combinatorial Auctions Using Rule-Based Bids, Decision Support Systems 34: 59–74.
207
Jones, J.L., and Koehler, G.J. (2005), A Heuristic for Winner Determination in Rule-Based Combinatorial Auctions, INFORMS Journal on Computing 17(4): 475–489.
Kagel, J.H. (1995), Auctions: A Survey of Experimental Research, in: Kagel, J.H., and Roth, A.E. (eds.), Handbook of Experimental Economics, Princeton, NJ, Princeton University Press.
Kagel, J.H., Harstad, R.M., and Levin, D. (1987), Information Impact and Allocation Rules in Auctions with Affiliated Private Values: A Laboratory Study, Econometrica 55(6): 1275–1304.
Kagel, J.H., and Levin, D. (1986), The Winner’s Curse and Public Information in Common Value Auctions, American Economic Review 76: 894–920.
Keeney, R., and Raiffa, H. (1976), Decisions with Multiple Objectives: Preferences and Value Tradeoffs, Wiley, New York.
Kelly, F., and Steinberg, R. (2000), A Combinatorial Auction with Multiple Winners for Universal Service, Management Science 46(4): 586–596.
Klemperer, P. (1998), Auctions with almost common values: the ‘Wallet Game’ and its applications, European Economic Review 42: 757–769.
Klemperer, P. (2002), What Really Matters in Auction Design, The Journal of Economic Perspectives 16(1): 169–189.
Koppius, O., and van Heck, E. (2002), Information Architecture and Electronic Market Performance in Multidimensional Auctions, working paper, Rotterdam School of Management, Erasmus University Rotterdam.
Korhonen, P., and Wallenius, J. (1996), Letter to the Editor: Behavioral Issues in MCDM: Neglected Research Questions, Journal of Multi-Criteria Decision Analysis 5: 178–182.
Krishna, V. (2002), Auction Theory, Academic Press, San Diego.
Kumar, V., Kumar, S., Tiwari, M.K., and Chan, F.T.S. (2006), Auction-Based Approach to Resolve the Scheduling Problem in the Steel Making Process, International Journal of Production Research 44: 1503–1522.
Kwasnica, A.M., Ledyard, J.O., Porter, D., and C. DeMartini (2005), A New and Improved Design for Multiobject Iterative Auctions, Management Science 51: 419–434.
Kwon, R.H., Anandalingam, G., and Ungar, L.H. (2005), Iterative Combinatorial Auctions with Bidder-Determined Combinations, Management Science 51(3): 407–418.
Köksalan, M, Leskelä, R.-L., Wallenius, H., and Wallenius, J. (2009), Improving Efficiency in Multiple-Unit Combinatorial Auctions: Bundling Bids from Multiple Bidders, forthcoming in Decision Support Systems.
Larichev, O. (1984), Psychological Validation of Decision Methods, Journal of Applied Systems Analysis 11: 37–46.
208
Ledyard, J.O., Olson, M., Porter, D., Swanson, J.A., and Torma, D.P. (2002), The First Use of a Combined-Value Auction for Transportation Services, Interfaces 32(5): 4–12.
Ledyard, J.O., Porter, C., and Rangel, A. (1997), Experiments Testing Multiobject Allocation Mechanisms, Journal of Economics & Management Strategy 6(3): 639–675.
Leskelä, R.-L., Teich, J., Wallenius, H., and Wallenius, J. (2007), Decision Support for Multi-Unit Combinatorial Auctions, Decision Support Systems 43: 420–434.
Levin, D., Kagel, J.H., and Richard, J.-F. (1996), Revenue Effects and Information Processing in English Common Value Auctions, American Economic Review 86(3): 442–460.
Lucking-Reiley, D., (1999), Using Field Experiments to Test Equivalence Between Auction Formats: Magic on the Internet, American Economic Review 89(5): 1063–1080.
Maasland, E., and Onderstal, S. (2006), Going, Going, Gone! A Swift Tour of Auction Theory and Its Applications, De Economist 154: 197–249.
Maskin, E., and Riley, J. (1989), Optimal Multi-Unit Auctions, in: Hahn, F. (ed.), The Economics of Missing Markets and Information, Oxford University Press, Oxford, UK.
McAfee, R., and McMillan, J. (1987), Auctions and Bidding, Journal of Economic Literature 25(2): 699–738.
McAfee, R., and McMillan, J. (1996), Analyzing the Airwaves Auction, Journal of Economic Perspectives 10(1): 159–175.
McMillan, J. (1994), Selling Spectrum Rights, Journal of Economic Perspectives 8(3): 145–162.
Metty, T., Harlan, R., Samelson, Q., Moore, T., Morris, T., Sorensen, R., Schneur, A., Raskina, O., Schneur, R., Kanner, J., Potts, K., and Robbins, J. (2005), Reinventing the Supplier Negotiation Process at Motorola, Interfaces 35: 7–23.
Milgrom, P.R. (1989), Auctions and Bidding: A Primer, Journal of Economic Perspectives 3(3): 3–22.
Milgrom, P. (1998), Game Theory and the Spectrum Auctions, European Economic Review 42: 771–778.
Milgrom, P. (2000), Putting Auction Tehory to Work: The Simultaneous Ascending Auction, The Journal of Political Economy 108(2): 245–272.
Milgrom, P.R., and Weber, R.J. (1982), A Theory of Auctions and Competitive Bidding, Econometrica 50(5): 1089–1122.
Mishra, D., and Parkes, D.C. (2007), Ascending price Vickrey auctions for general valuations, Journal of Economic Theory 132: 335–366.
Mishra, D., and Veeramani, D. (2007), Vickrey-Dutch procurement auction for multiple items, European Journal of Operational Research 180: 617–629.
209
Mito, M. and Fujita, S. (2004), On Heuristics for Solving Winner Determination Problem in Combinatorial Auctions, Journal of Heuristics 10: 507–523.
Murray, J.D., and White, R.W. (1983), Economies of Scale and Economies of Scope in Multiproduct Financial Institutions: A Study of British Columbia Credit Unions, The Journal of Finance 38(3): 887–902.
Myerson, R.B. (1981), Optimal Auction Design, Mathematics of the Operations Research 6(1): 58–73.
Nisan, N. (2000), Bidding and Allocation in Combinatorial Auctions, Proceedings of ACM Conference on Electronic Commerce (EC 00), ACM Press, New York: 1–12.
Nisan, N. (2006), Bidding Languages for Combinatorial Auctions, Chapter 9 in: P. Cramton, Y. Shoham, and R. Steinberg (eds.), Combinatorial Auctions, MIT Press: 215–232.
Ono, C., Nishiyama, S., and Horiuchi, H. (2003), Reducing complexity in winner determination for combinatorial ascending auctions, Electronic Commerce Research and Applications 2: 176–186.
Papadimitriou, C.H., and Steiglitz, K. (1982), Combinatorial Optimization: Algorithms and Complexity, Prentice Hall, New Jersey.
Park, S., and Rothkopf, M. (2005), Auctions with Bidder-Determined Allowable Combinations, European Journal of Operational Research 161(2): 399–415.
Parkes, D.C. (1999), iBundle: An Efficient Ascending Price Bundle Auction, Proceedings of First ACM Conference on Electronic Commerce (EC-99), Denver, Colorado, 1999: 148–157.
Parkes, D.C. (2001), Iterative Combinatorial Auctions: Achieving Economic and Computational Efficiency, Doctoral dissertation, Computer and Information Science, University of Pennsylvania, Philadelphia.
Parkes, D.C. (2006), Iterative Combinatorial Auctions, Chapter 2 in: P. Cramton, Y. Shoham, and R. Steinberg (eds.), Combinatorial Auctions, MIT Press: 41–77.
Parkes, D., and Ungar, L.H. (2000), Iterative Combinatorial Auctions: Theory and Practice, Proceedings of the 17th National Conference on Artificial Intelligence (AAAI-00): 74–81.
Pekeč, A., and Rothkopf, M. (2000), Making the FCC’s First Combinatorial Auction Work Well, Comment on FCC Public Notice, DA 00-1075, http://wireless.fcc.gov/auctions/31/releases/workwell.pdf, accessed Nov 30th, 2008.
Pekeč, A., and Rothkopf, M. (2003), Combinatorial Auction Design, Management Science 49: 1485–1503.
Pinker, E.J., Seidmann, A., and Vakrat, Y. (2003), Managing Online Auctions: Current Business and Research Issues, Management Science 49(11): 1457–1484.
Porter, D., Rassenti, S., Roopnarine, A., and Smith, V. (2003), Combinatorial Auction Design, Proceedings of the National Academy of Sciences of the United States of America (PNAS), 100(19): 11153–11157.
210
Puro, L. (2009), Strategic Decision Making in Online Auctions: Prosper.com People-to-People Lending Site, Master’s Thesis, Helsinki University of Technology, Dept. of Industrial Engineering and Management.
Rassenti, S.J., Smith, V.L., and Bulfin, R.L. (1982), A Combinatorial Auction Mechanism for Airport Time Slot Allocation, The Bell Journal of Economics 13(2): 402–417.
Riley, J.G., and Samuelson, W.F. (1981), Optimal Auctions, The American Economic Review 71(3):381–392.
Robinson, M.S. (1985), Collusion and the Choice of Auction, RAND Journal of Economics 16: 141–145.
Roth, A.E. and Ockenfels, A. (2002), Last-Minute Bidding and the Rules for Ending Second-Price Auctions: Evidence from eBay and Amazon Auctions on the Internet, The American Economic Review 92(4):1093–1103.
Rothkopf, M., and Harstad, R. (1994), Modeling Competitive Bidding: A Critical Essay, Management Science 40(3): 364–384.
Rothkopf, M., Pekeč, A., and Harstad, R. (1998), Computationally Manageable Combinatorial Auctions, Management Science 44(8): 1131–1147.
Sandholm, T. (2000), Approaches to Winner Determination in Combinatorial Auctions, Decision Support Systems 28(1-2): 165–176.
Sandholm, T. (2002), Algorithm for Optimal Winner Determination in Combinatorial Auctions, Artificial Intelligence 135: 1–54.
Sandholm, T. (2007), Expressive Commerce and Its Application to Sourcing: How We Conducted $35 Billion of Generalized Combinatorial Auctions, AI Magazine 28(3): 45–58.
Sandholm, T. and Boutlier, C. (2006), Preference Elicitation in Combinatorial Auctions, Chapter 10 in: P. Cramton, Y. Shoham, and R. Steinberg (eds.), Combinatorial Auctions, MIT Press: 233–263.
Sandholm, T., Levine, D., Concordia, M., Martyn, P., Hughes, R., Jacobs, J., and Begg, D. (2006), Changing the Game in Strategic Sourcing at Procter & Gamble: Expressive Competition Enabled by Optimization, Interfaces 36(1): 55–68.
Sandholm, T., and Suri, S. (2003), BOB: Improved Winner Determination in Combinatorial Auctions and Generalizations, Artificial Intelligence 145: 33–58.
Sandholm, T., and Suri, S. (2006), Side Constraints and Non-Price Attributes in Markets, Games and Economic Behavior 55: 321–330.
Sandholm, T., Suri, S., Gilpin, A., and Levine, D. (2005), CABOB: A Fast Optimal Algorithm for Winner Determination in Combinatorial Auctions, Management Science 51(3): 374–390.
Schnizler, B., Neumann, D., Veit, D., and Weinhardt, C. (2008), Trading Grid services – a multi-attribute combinatorial approach, European Journal of Operational Research 187: 943–961.
211
Shah, H.S., Joshi, N.R., Sureka, A., and Wurman, P.R. (2003), Mining eBay: Bidding Strategies and Shill Detection, in: Zaïane, O.R., Srivastava, J., Spiliopoulou, M., and Masand, B. (eds.), WEBKDD 2002 – Mining Web Data for Discovering Usage Patterns and Profiles, Lecture Notes in Artificial Intelligence 2703: 17–34.
Sheffi, Y. (2004), Combinatorial Auctions in the Procurement of Transportation Services, Interfaces 34: 245–252.
Simon, H. (1955), A Behavioral Model of Rational Choice, Quarterly Journal of Economics 69: 7–19.
Srivinas, W., Tiwari, M.K., and Allada, V. (2004), Solving the Machine-Loading Problem in a Flexible Manufacturing System Using a Combinatorial Auction-Based Approach, International Journal of Production Research 42(9): 1879–1893.
Teich, J.E., Wallenius, H., and Wallenius, J. (1999), Multiple Issue Auction and Market Algorithms for the World Wide Web, Decision Support Systems 26(1): 49–66.
Teich, J.E., Wallenius, H., Wallenius, J., and Koppius, O.R. (2004), Emerging Multiple Issue E-auctions, European Journal of Operational Research 159: 1–16.
Teich, J.E., Wallenius, H., Wallenius, J., and Zaitsev, A. (2001), Designing Electronic Auctions: An Internet-Based Hybrid Procedure Combining Aspects of Negotiations and Auctions, Electronic Commerce Research 1: 301–314.
Teich, J.E., Wallenius, H., Wallenius, J., and Zaitsev, A. (2006), A Multi-Attribute e-Auction Mechanism for Procurement: Theoretical Foundations, European Journal of Operational Research 175: 90–100.
Tenorio, R. (1999), Multiple Unit Auctions with Strategic Price-Quantity Decisions, Economic Theory 13: 247–260.
Tukiainen, J. (2008), Testing for common cost in the City of Helsinki bus transit auctions, International Journal of Industrial Organization 26: 1308–1322.
Vakrat, Y. and Seidmann, A., 2000, Implications of the bidders’ arrival process on the design of online auctions, Proceedings of the 33rd Hawaii International Conference on System Sciences, 2000: 1–10.
Vickrey, W. (1961), Counterspeculation, Auctions, and Competitive Sealed Tenders, The Journal of Finance 16(1): 8–37.
Vohra, R., and Weber, R.J. (2000), A ‘Bids and Offers’ Approach to Package Bidding, Comment on FCC Public Notice, DA 00-1075, http://wireless.fcc.gov/auctions/31/releases/bid_offr.pdf, accessed Nov 30th, 2008.
Wilson, R.B. (1969), Competitive Bidding with Disparate Information, Management Science 15(7): 446–448.
Wilson, R.B. (1977), A Bidding Model of Perfect Competition, The Review of Economic Studies 44(3): 511–518.
Wilson, R.B. (1979), Auctions of Shares, Quarterly Journal of Economics 93(4): 557–585.
212
Wilson, R.B. (1992), Strategic Analysis of Auctions, in: Aumann, R.J., and Hart, S. (eds.), Handbook of Game Theory Vol. I, North-Holland, Amsterdam: 227–271.
Wurman, P., and Wellman, M. (2000), AkBA: A Progressive, Anonymous-Price Combinatorial Auction, Proceedings of Second ACM Conference on Electronic Commerce (EC-00), Minneapolis, Minnesota, 2000: 21–29.
Yang, S., Segre, A.M., and Codenotti, B. (2009), An optimal multiprocessor combinatorial auction solver, Computers and Operations Research 36(1): 149–166.
Yokoo, M., Sakurai, Y. and Matsubara, S. (2004), The effect of false-name bids in combinatorial auctions: new fraud in internet auctions, Games and Economic Behavior 46: 174–188.
Zionts, S., and Wallenius, J. (1976), An Interactive Programming Method for Solving the Multiple Criteria Problem, Management Science 22(6): 652–663.
Özer, A.H., and Özturan, C. (2009), A model and heuristic algorithms for multi-unit nondiscriminatory combinatorial auction, Computers & Operations Research 36: 196–208.
213
APPENDIX 1: FIXED COST PARAMETERS IN THE FIRST
SIMULATION STUDY
Table 35 Upper and lower limits for the uniform distributions from which the fixed cost
parameters were drawn in the first simulation study
Lower Upper Lower Upper
F1 1000 1200 8000 10000
F2 1000 1200 7000 11000
F3 700 1000 5000 7000
F4 1500 2000 9000 12000
F5 1600 2500 10000 13000
F12 1700 2300 13000 18000
F13 1700 2200 12000 16000
F14 2100 3000 17000 20000
F15 2500 3500 18000 22000
F23 1600 2100 11000 15000
F24 2500 3000 16000 23000
F25 2500 4000 17000 24000
F34 2200 3000 13000 16000
F35 2000 3000 15000 20000
F45 2900 4000 17000 22000
F123 2300 3200 18000 24000
F124 3100 4200 21000 27000
F125 3200 5000 22000 30000
F134 3200 4200 20000 26000
F135 3300 4600 22000 29000
F145 3600 5200 23000 29000
F234 3000 4000 23000 27000
F235 3100 4500 24000 30000
F245 3800 5400 25000 32000
F345 3700 5500 22000 26000
F1234 4100 5500 30000 38000
F1235 3900 5400 31000 40000
F1245 4600 6300 34000 41000
F1345 4500 6200 28000 38000
F2345 4600 6200 34000 40000
F12345 5500 7500 42000 50000
"Low" Fixed Cost "High" Fixed Cost
Fijk = the fixed cost of producing items i, j and k jointly
214
APPENDIX 2: FORMULATION OF THE QSP WITH TRUE COST
FUNCTIONS
Here I will present the formulation of the QSP for the case of ten bidders and five
items, which corresponds to the one of the settings in the first simulation study. The
formulation in the case of three items is analogous but simpler due to fewer cost
function parameters. The formulation is presented in the context of the simulation
study instead of in a general form in order to make it easier to read. The QSP is
formulated for bidder m. To simplify the notation, the index indicating bidder m is
omitted from the cost function variables and parameters.
The objective is to find a bid price pnew, bid quantities qnew,kj and the combination of
complementing bids from other bidders that maximizes the profit for the bidder. The
status of bidder i’s jth bid is indicated by xij (xij =1 indicates active status, xij = 0 indicates
inactive status). Let Fijk denote bidder m’s fixed cost of producing items i, j, and k
jointly, and ck the variable cost of item k. Because of the discontinuous cost functions,
the formulation becomes quite cumbersome and lengthy. Auxiliary variables yi, yij, yijk,
yijkl and yijklm ∈{0,1} ensure that the correct fixed cost parameter is applied. E.g. if in the
optimal bid, items 1, 3, 4 and 5 assume a positive value, the system should set y1345 = 1,
and others to zero. For the construction of the constraints that ensures that this in fact
happens, we need several (mutually exclusive) variables for the quantities of each item.
These variables are denoted qnew,kj, where k indicates the item (k = 1, … , 5), and j the
different variables for the same item (j = 1, … , 16). In other words, there is one
quantity variable for each combination that item k is a part of. Thus, for any item k
only one qnew,kj can assume a positive value.
The problem is to maximize:
123451234523452345
3
2
451451
5
4
123123
3
1
4545
5
4
2323
5
4
1313
5
3
1212
4545
5
4
33
5
3
22
5
2
11
16
1
5
1
5
1
,
yFyF
yFyFyFyFyFyF
yFyFyFyFyFqcp
k
kk
k
kk
k
kk
k
kk
k
kk
k
kk
k
kk
k
kk
k
kk
j k
kk
k
kjnewknew
−−
−−−−−−
−−−−−−
∑∑∑∑∑∑
∑∑∑∑ ∑∑
======
==== ==
(51)
215
subject to the constraints:
(i) Total cost to the buyer must decline 5% from the current best *C :
*10
1
95.0 Cpxp new
i
ijij ≤+∑=
(52)
Notice that in the first simulation study we are considering the point in time in the
auction when there is only one bid from each bidder i in the existing bid stream (j = 1),
so there is no need to sum over the j’s.
(ii) The buyer’s demand for each item (= 1000 units) must be fulfilled, qijk is the
quantity of good k in bidder i’s initial bid (again, no need to sum over j):
5,...,1 100016
1
,
10
1
=∀≥+∑∑==
kqxqj
kjnew
i
ijijk (53)
(iii) Exactly one fixed cost is chosen:
1123452345
3
2
451
5
4
123
3
1
45
5
4
23
5
3
5
4
131245
5
4
3
5
3
2
5
2
1
5
1
=++++
++++++++
∑∑
∑∑∑ ∑∑∑∑∑
==
=== =====
yyyy
yyyyyyyyy
k
k
k
k
k
k
k
k
k k
kk
k
k
k
k
k
k
k
k
(54)
(iv) The correct fixed cost should be chosen
The constraint (iii) assures that only one of constraints (iv) can be nonbinding. The
objective function (maximization) assures that given the bid quantities, the minimum
fixed cost is chosen. E.g. if the bid quantities are nonzero for only items 1 and 2, the
algorithm will set y12 = 1 allowing qnew,1,2 and qnew,2,2 assume a positive value. The
constraints would allow the algorithm to set for instance y12345 = 1 (qnew,1,16 and qnew,2,16
would assume positive values), but because F12 < F12345 it would not be optimal.
216
5,...,101,, =∀≤− kMyq kknew
0
0
0
0
0
0
0
0
0
0
455,5,5,4,
354,5,5,3,
344,4,4,3,
253,5,5,2,
243,4,4,2,
233,3,3,2,
152,5,5,1,
142,4,4,1,
132,3,3,1,
122,2,2,1,
≤−+
≤−+
≤−+
≤−+
≤−+
≤−+
≤−+
≤−+
≤−+
≤−+
Myqq
Myqq
Myqq
Myqq
Myqq
Myqq
Myqq
Myqq
Myqq
Myqq
newnew
newnew
newnew
newnew
newnew
newnew
newnew
newnew
newnew
newnew
0
0
0
0
0
0
0
0
0
0
34511,5,11,4,11,3,
24510,5,10,4,11,2,
2359,5,10,3,10,2,
2349,4,9,3,9,2,
1458,5,8,4,11,1,
1357,5,8,3,10,1,
1347,4,7,3,9,1,
1256,5,8,2,8,1,
1246,4,7,2,7,1,
1236,3,6,2,6,1,
≤−++
≤−++
≤−++
≤−++
≤−++
≤−++
≤−++
≤−++
≤−++
≤−++
Myqqq
Myqqq
Myqqq
Myqqq
Myqqq
Myqqq
Myqqq
Myqqq
Myqqq
Myqqq
newnewnew
newnewnew
newnewnew
newnewnew
newnewnew
newnewnew
newnewnew
newnewnew
newnewnew
newnewnew
0
0
0
0
0
234515,5,15,5,15,3,15,2,
134514,5,14,4,14,3,15,1,
124513,5,13,4,14,2,14,1,
123512,5,13,3,13,2,13,1,
123412,4,12,3,12,2,12,1,
≤−+++
≤−+++
≤−+++
≤−+++
≤−+++
Myqqqq
Myqqqq
Myqqqq
Myqqqq
Myqqqq
newnewnewnew
newnewnewnew
newnewnewnew
newnewnewnew
newnewnewnew
01234515,5,16,4,16,3,16,2,16,1, ≤−++++ Myqqqqq newnewnewnewnew
(55)
(v) Capacity constraints as defined in the first simulation study:
16,...1,5,...,1500, ==∀≤ jkq kjnew (56)
217
(vi) Only one bid can be active per bidder:
01 =mx (57)
Bidder m’s new bid will become active, so her initial bid cannot be active. There is
only one bid from all the other bidders, at this point in time, hence there is no need for
additional constraints yet.
(vii) The bid status variables of the initial bids are binary variables:
{ } 10,...,11,0 =∀∈ ixij (58)
(viii) Every yijklm is binary:
{ }
{ }
{ }
{ }
{ }1,0
5,4,5,...,3,5,...,2,5,...,11,0
5,...,3,5,...,2,5,...,11,0
5,...,2,5,...,11,0
5,...,11,0
12345 ∈
====∀∈
===∀∈
==∀∈
=∀∈
y
lkjiy
kjiy
jiy
iy
ijkl
ijk
ij
i
(59)
218
APPENDIX 3: FORMULATION OF THE EFFICIENT ALLOCATION
PROBLEM
The efficient allocation problem in the second simulation study is in some respects
similar to the formulation of the quantity support problem with true cost functions
presented in Appendix 2. The logic by which the correct fixed cost is chosen is the
same. Also the constraints assuring that buyer’s demand is fulfilled, and that only one
bid per bidder can be accepted are the same. The objective function is different, but it
also contains similar elements. The efficient allocation problem is formulated for the
case of five items and 15 bidders, which corresponds to one of the simulation designs.
A general formulation of the efficient allocation problem would be much more
difficult for readers to follow.
Let Fjklmn denote the fixed cost of producing items j, k, l, m and n jointly, and cij bidder
i’s variable cost of producing item j. The capacity constraints for bidder i’s item j are
denoted aij. The objective is to select the combination of bids from different bidders
that minimize the total production cost. Because of the discontinuous cost functions,
the formulation becomes quite cumbersome and lengthy. Auxiliary variables yi,j, yi,jk,
yi,jkl, yi,jklm and yi,jklmn { }1,0∈ ensure that the correct fixed cost parameter is applied. E.g. if
in the optimal allocation, items 1, 3, 4 and 5 assume a positive value for bidder 1, the
system should set y1,1345 = 1, and other yi’s to zero. For the construction of the constraints
that ensures that this in fact happens, we need several (mutually exclusive) variables for
the quantities of each item. These variables are denoted qijk, where i indicates the
bidder (i = 1, … , 15), j the item (j = 1, … , 5) and k enumerates the different variables
for the same item (k = 1, … , 16). For any bidder i and item j only one qi,j,k can assume
a positive value.
219
The problem is to minimize:
)
(
12345,12345,2345,2345,
3
2
451,451,
5
4
123,123,
3
1
45,45,
5
4
23,23,
5
4
13,13,
5
3
12,12,
45,45,
5
4
3,3,
5
3
2,2,
5
2
1,1,
5
1
5
1
,,
16
1
15
1
iiii
m
mimi
m
mimi
m
mimi
m
mimi
m
mimi
m
mimi
ii
m
mimi
m
mimi
m
mimi
j m
mimi
k
ijkij
i
yFyFyF
yFyFyFyFyF
yFyFyFyFyFqc
+++
+++++
+++++
∑
∑∑∑∑∑
∑∑∑∑ ∑∑∑
=
=====
==== ===
(60)
subject to the constraints:
(i) The buyer’s demand for each item (= 600 units) must be fulfilled
5,...,160015
1
16
1
=∀≥∑∑= =
jqi k
ijk (61)
(ii) At most one fixed cost is chosen for each bidder
15,...,1112345,2345,
3
2
451,
5
4
123,
3
1
45,
5
4
23,
5
4
13,
5
3
12,45,
5
4
3,
5
3
2,
5
2
1,
5
1
,
=∀<+++++
+++++++
∑∑∑
∑∑∑∑∑∑∑
===
=======
iyyyyy
yyyyyyyy
ii
m
mi
m
mi
m
mi
m
mi
m
mi
m
mii
m
mi
m
mi
m
mi
m
mi
(62)
(iii) The correct fixed cost should be chosen for each bidder i = 1, … , 15. The
following constraints allow a yi,jklmn to assume the value of one only if the right
combination of item quantities assumes a positive value.
5,...,10,1,, =∀<− jMyq jiji
0
0
0
0
0
0
0
0
0
0
45,5,5,5,4,
35,4,5,5,3,
34,4,4,4,3,
25,3,5,5,2,
24,3,4,4,2,
23,3,3,3,2,
15,2,5,5,1,
14,2,4,4,1,
13,2,3,3,1,
12,2,2,2,1,
≤−+
≤−+
≤−+
≤−+
≤−+
≤−+
≤−+
≤−+
≤−+
≤−+
iii
iii
iii
iii
iii
iii
iii
iii
iii
iii
Myqq
Myqq
Myqq
Myqq
Myqq
Myqq
Myqq
Myqq
Myqq
Myqq
(63)
220
0
0
0
0
0
0
0
0
0
0
345,11,5,11,4,11,3,
245,10,5,10,4,11,2,
235,9,5,10,3,10,2,
234,9,4,9,3,9,2,
145,8,5,8,4,11,1,
135,7,5,8,3,10,1,
134,7,4,7,3,9,1,
125,6,5,8,2,8,1,
124,6,4,7,2,7,1,
123,6,3,6,2,6,1,
≤−++
≤−++
≤−++
≤−++
≤−++
≤−++
≤−++
≤−++
≤−++
≤−++
iiii
iiii
iiii
iiii
iiii
iiii
iiii
iiii
iiii
iiii
Myqqq
Myqqq
Myqqq
Myqqq
Myqqq
Myqqq
Myqqq
Myqqq
Myqqq
Myqqq
0
0
0
0
0
0
12345,16,5,16,4,16,3,16,2,16,1,
2345,15,5,15,4,15,3,15,2,
1345,14,5,14,4,14,3,15,1,
1245,13,513,4,14,2,14,1,
1235,12,5,13,3,13,2,13,1,
1234,12,4,12,3,12,2,12,1,
≤−++++
≤−+++
≤−+++
≤−+++
≤−+++
≤−+++
iiiiii
iiiii
iiiii
iiii
iiiii
iiiii
Myqqqqq
Myqqqq
Myqqqq
Myqqqq
Myqqqq
Myqqqq
15,...,1=∀ i
The constraints above do not rule out the possibility that e.g. yi,12345 = 1, even though
only four items or less actually assume positive quantities. However, since Fi,12345 is set to
be larger than any other fixed cost, the objective to minimize total cost will choose the
lowest fixed cost allowed by the constraints. The same argument applies to any other Fi
as well.
(iv) Capacity constraints
16,...,1,5,...,1,15,...,1 ===∀≤ kjiaq ijijk (64)
(v) Every auxiliary variable is binary
{ }
{ }
{ }
{ }
{ } 15,...,11,0
5,4,5,...,3,5,...,2,5,...,1,15,...,11,0
5,...,3,5,...,2,5,...,1,15,...,11,0
5,...,2,5,...,1,5,...,11,0
5,...,1,15,...,11,0
12345,
,
,
,
,
=∀∈
=====∀∈
====∀∈
===∀∈
==∀∈
iy
mlkjiy
lkjiy
kjiy
jiy
i
jklmi
jkli
jki
ji
(65)
221
APPENDIX 4: BIDDERS’ COST FUNCTIONS IN THE LABORATORY
EXPERIMENTS
Table 36 presents the cost function parameters for bidders in equal capacities (A)
auctions. In groups of five students, the cost functions used were those of Bidders 5, 8,
10, 12 and 15. Bidder 1 was added to the six person group, and Bidder 11 to the seven
person groups. The capacities are equal (300, 300, 300, 300, 300) for all bidders.
Table 37 presents the cost function parameters for bidders in unequal capacities (B)
auctions. In the groups with only six bidders, Bidder 9 was removed from the auction.
The bidders’ maximum capacities are presented in Table 38.
222
Table 36 Bidders’ cost function parameters in the equal capacities (A) auctions
Bidder 1 Bidder 5 Bidder 8 Bidder 10 Bidder 11 Bidder 12 Bidder 15
c1 59,38 63,47 61,53 56,31 65,53 53,51 55,14
c2 56,64 57,93 56,72 66,59 55,12 63,85 54,03
c3 66,51 59,30 58,00 65,03 66,23 65,32 56,86
c4 60,85 60,21 61,50 61,67 61,59 53,36 63,43
c5 56,27 55,43 57,83 56,15 61,87 60,21 59,14
F1 16644 19329 19673 19523 17690 17823 19984
F2 18364 18454 19960 17703 16585 18183 16335
F3 17773 19657 19426 17371 17675 16475 17316
F4 16205 16037 17720 16041 16948 17606 17746
F5 18027 17700 16883 17122 18696 18393 19308
F12 26753 24299 27610 28980 24416 25395 24408
F13 24237 24180 28605 29348 26447 26261 29530
F14 26535 24491 27360 29895 26915 24002 25775
F15 29568 24169 27738 26707 26455 29454 28900
F23 26655 26067 29387 25488 24119 29981 24804
F24 27489 29978 27523 29698 28452 29236 26080
F25 27329 25960 26733 26804 27111 25219 27939
F34 28784 27738 26447 24977 24157 29108 28124
F35 26035 24982 26579 27734 29911 27822 24020
F45 25853 25066 26335 27644 27911 28575 25683
F123 32027 33187 33649 39980 38591 38147 35436
F124 38542 33124 35341 37485 38648 34275 34703
F125 35553 38147 32306 39640 34664 34620 34437
F134 36471 33870 37599 34281 33366 33626 35147
F135 35695 32205 32983 37035 33573 35710 32926
F145 38652 32273 37331 34045 33142 34218 36823
F234 34789 37650 32590 36189 37196 33867 38993
F235 33640 37321 39511 38119 33195 36955 34543
F245 34385 33144 38243 33370 34933 39823 33217
F345 38968 37475 35569 34928 37218 38972 39675
F1234 47512 42397 42200 41194 48537 41679 47694
F1235 43276 49508 43240 42877 47978 42148 49835
F1245 49030 43952 42911 44056 42098 40240 40949
F1345 42334 48547 49552 48188 40433 42921 49354
F2345 48607 49466 48132 43639 46775 41971 40184
F12345 54550 57982 56416 50253 50363 50539 50770
Vari
ab
le C
os
tsF
ixed
Co
sts
223
Table 37 Bidders' cost function parameters in the unequal capacities (B) auctions
Bidder 1 Bidder 3 Bidder 4 Bidder 8 Bidder 9 Bidder 13 Bidder 15
c1 58,06 64,20 62,50 57,76 53,85 64,03 56,42
c2 54,92 53,61 57,95 66,27 55,01 65,53 65,93
c3 53,88 56,35 66,62 62,05 53,61 62,37 62,51
c4 65,52 59,30 56,61 56,37 65,19 54,31 63,24
c5 55,51 59,19 56,11 54,46 55,25 64,72 65,84
F1 19635 17355 18239 16258 17196 16262 18466
F2 16576 18505 17497 19397 19719 19580 19094
F3 19135 17729 19955 17188 18652 16345 18284
F4 18093 18624 16583 16041 19937 18950 18681
F5 19704 16071 19752 19516 18491 16676 17829
F12 24964 27114 28477 26161 28601 28430 26739
F13 26764 26657 26962 24764 24926 25063 27071
F14 25847 25108 29133 27139 26637 25105 25626
F15 25042 28210 26492 24684 24773 27615 25549
F23 27501 29891 28301 29100 25777 26823 25422
F24 28442 24164 29082 24868 24889 24857 27556
F25 26582 26570 28532 24972 25423 25184 28519
F34 28385 29860 29385 27243 26194 26385 25532
F35 25112 24656 26220 26547 24586 27173 27953
F45 26151 28130 27062 28896 26993 28669 27715
F123 38070 33185 39646 36258 35136 36352 38739
F124 35861 36397 32763 39856 35548 34976 36750
F125 38723 37323 33967 34399 34256 36813 39910
F134 33403 36506 37369 38713 36166 37043 38794
F135 35990 36207 34658 33096 33748 32624 35264
F145 33484 39025 38708 36367 34471 37745 36521
F234 33749 36226 33008 38549 36967 32360 35033
F235 32767 34727 36100 38136 39249 33489 39010
F245 36642 38103 35639 33369 35321 32031 35594
F345 35093 32670 32373 39741 33992 39524 38494
F1234 49256 48888 47454 43536 47339 47114 42418
F1235 41993 44341 49214 49981 47720 49611 41734
F1245 45501 42292 48059 47062 49374 46941 47954
F1345 42764 41781 49930 40662 46926 41533 43732
F2345 48579 46659 45918 41613 49433 42818 41537
F12345 55647 54080 52026 57525 59721 56902 49149
Vari
ab
le C
os
tsF
ixed
Co
sts
Table 38 Bidders' capacities in the unequal capacities (B) auctions
Bidder 1 Bidder 3 Bidder 4 Bidder 8 Bidder 9 Bidder 13 Bidder 15
Item 1 225 300 300 300 300 150 300
Item 2 225 300 150 300 300 300 300
Item 3 225 300 300 150 225 300 150
Item 4 225 150 300 300 150 300 300
Item 5 300 150 300 300 225 300 150
224
APPENDIX 5: AUCTION EXPERIMENT PARTICIPANT
QUESTIONNAIRE
NAME: _______________
STUDENT ID: _________
1. Have you participated in any kind of online auctions before (yes/no)?
2. Did you place bids only with the help of the support tools after it became available
(yes/no)?
Answer the following questions by choosing the appropriate number from 0-5.
0 = I don’t know
1 = I totally disagree
2 = I somewhat disagree
3 = I don’t agree or disagree
4 = I somewhat agree
5 = I totally agree
3. I understood the rules of the auction game 0 1 2 3 4 5
4. I understood what my goal was in the game 0 1 2 3 4 5
5. The CombiAuction site was easy to use 0 1 2 3 4 5
6. I understood how the price support tool works 0 1 2 3 4 5
7. I understood how the quantity support tool works 0 1 2 3 4 5
8. The price support tool was easy to use 0 1 2 3 4 5
9. The quantity support tool was easy to use 0 1 2 3 4 5
10. The price support tool was helpful 0 1 2 3 4 5
11. The quantity support tool was helpful 0 1 2 3 4 5
12. Please describe your bidding strategy during the two auctions (When did you bid?
How often did you bid? How did you choose the items, quantities and price in your
bids when you didn’t use the support tools? Etc.).
13. How did your experience in the A auction affect your bidding in the B auction?
14. Did you monitor the auctions within the 10 minutes before it was scheduled to
close? If you did not have active bids, did you use the support tools? Were you
successful in finding profitable bids?
15. How would you improve the usability of the CombiAuction system?
16. Is there anything else you want to comment on about the game?